Shahzad Bhatti

August 19, 2020

Review of “Building Secure and Reliable Systems”

Filed under: Computing,Technology — admin @ 5:33 pm

The “Building Secure and Reliable Systems” book shares best practices from Google’s security and SRE engineers. Here is a summary of these best practices:

The first chapter discusses tradeoff between security and reliability, e.g. reliability protects against non-malicious failure but may expand security surface via redundancy whereas security risk comes from adversarial attacks. Both reliability and security need confidentiality, integrity and availability but with different perspectives. Complex systems are difficult to reason so you must apply “Defense in depth”, “Principle of least privilege” and “Distinct failure domains” to limit the blast radius of failure. For example, Google uses geographic regions to limit the scope of credentials.

The second chapter focuses on “security adversaries” and “attack motives” who may come from hobbyists, hacktivist, researchers, criminals, cyber warfare, insiders and other background. You can apply CAPTCHA, automation/AI, zero trust, multi-party authorization, auditing/detection and recoverability to protect against these attacks.

The third chapter is part of second section of the book that focuses on designing secure and reliable systems. It introduces safe proxies in production environment that enforce authentication, multi-party authorization (MPA), auditing, rate limiting, zero touch, access control, etc. For example, Google uses CLI proxy to execute commands that are controlled via security policy, MPA and provides auditing/logs.

The chapter four examines security tradeoffs and reviews product features that may include functional and non-functional requirements (e.g. security, reliability, SLO dev velocity). Reliability and security are also considered emergent properties of system design and encompass entire product and services. The chapter also gives an example of design document template that includes sections for scalability, redundancy/reliability, dependencies, data-integrity, SLA, and security/privacy.

The chapter five discusses designing for least privilege that uses authentication and authorization. It also examines zero-trust networks that doesn’t grant any illegal access and zero-touch interfaces where all access is automated. It recommends writing small functions so that access control can be clearly defined, breaking glass in case of emergency to bypass certain authorization systems, auditing, testing for least privilege, multi-party authorization (MPA), three-factor authorization (3FA where access is approved from two platforms), business justifications, temporary access, proxies etc. This chapter also discusses tradeoffs of complex security with other factors such as company culture, data quality, user productivity, and development complexity.

The chapter six focuses on designing for understanding to reduce likelihood of security vulnerabilities and increase confidence in the system security. It defines system invariant, which is a property that is always true and can be used to assert security and reliability properties. It suggests using mental model to understand complex security system and explains identities, authentication, and access control concepts. When breaking a system into smaller components, the chapter recommends using trusted computing base (TCB) to create a security boundary that enforces security policies. In order to provide access from one TCB to another, you may issue end-user context ticket (EUC) that provides access temporarily. In order to standardize security policies, you may use a common framework for request dispatching, input sanitization, authentication, authorization, auditing, logging, monitoring, quota, load balancing, configuration, testing, dashboard, alerting, etc.

The chapter seven focuses on extensibility and new changes. For example, keeping dependencies up-to-date, automated testing, release frequently, using containers, micro services, etc.

The chapter eight focuses on resilience that describes the system’s ability to hold out against a major malfunction or disruption. It encourages designing the system with independent layers, modularization, redundancy, automation, security in defense, controlled degradation (partially failure), load shedding, throttling, automated response. You will need to consider tradeoffs between reliability and security, e.g. failing safe vs failing secure where reliability/safety may require ACL is “allow-all” but security may require ACL is “deny-all”. You can segment your network and Compartmentalize your system to reduce the blast radius. With micro-service architecture, you can assign distinct roles for each service and add geographic location or time as a scope of access. The chapter then defines failure domain, which is a type of blast radius control that creates isolation by partitioning a system into multiple equivalent but completely independent copies with its own data. Any of the individual partitions can take over for the entire system during an outage and help protect systems from global impact. You can validate the system continuously for failures using fuzzing and other types of testing.

The chapter nine discusses recoverability from random, accidental, software failures and errors. The chapter recommends designing emergency push system to simply be your regular push system turned up-to maximum for recovering it from failure. In order to prevent rollback to older-version, you can collect undesirable versions into a deny list or use white-list of allowed versions, which is used in the release system for verification. Also, you can maintain security version numbers (SVNs) and minimum acceptable security version numbers (MASVNs) and rotate signing keys, e.g.

ComponentState[DenyList] = ComponentState[DenyList].union(self[DenyList))
ComponentState[MASVN] = max(self[MASVN], ComponentState[MASVN])

def IsUpdateAllowed(self, Release, ComponentState, KeyDatabase):
  assert Release[Version] not in ComponentState[DenyList]
  assert Release[SVN] >= ComponentState[MASVN]
  assert VerifySignature(Release, KeyDatabase)

The chapter ten explains how to mitigate D.O.S. attacks where attacker may compromise vulnerable machines or launch amplification attacks. This chapter suggests using edge routers to throttle high-bandwidth attacks and eliminate attack traffic as early as possible. For example, You can use network and application load balancers to continually monitor incoming traffic. Other mitigating techniques include using caching proxies, minimize network requests (e.g. using spriting), minimize egress bandwidth, CAPTCHA, rate limit, monitoring/alerting (MTTD mean-time-to-detect, MTTR mean-time-to-repair), graceful degradation, exponential backoff, jitter, etc.

The chapter eleven is part of third section and focuses on maintaining trusted CA. For example, you can use secure and memory-safe languages to parse certificates or CSR requests. You may need to use third-party libraries but you can add testing for validation.

The chapter twelve focuses on writing code, e.g. using frameworks that enforce security and reliability. You can use RPC frameworks that may provide logging, authentication, authorization, rate-limiting. This chapter covers OWASP top vulnerabilities such as SQL injection that can be prevented by using parameterized SQLs; XSS that can be prevented by using sanitizing user input (safeHTML) and incremental rollout. Other coding techniques include simplicity, minimizing multi-level nesting/cyclometic complexity, eliminate yagni smells, pay tech-debt, refactoring. The chapter also suggests using memory-safe and strongly/static typed languages.

The chapter thirteen examines testing code using unit and integration tests. It also introduces other testing techniques such as fuzz testing, chaos engineering, static program analysis, code inspection tools (Error Prone for Java and Clang-Tidy), and formal methods.

The chapter fourteen describes deployment phase of software development that may include pushing the code, downloading a new binary, updating configuration, migrating database, etc. The chapter reviews threat model to prevent bad deployment such as accidental change, malicious change, bad configuration, stealing integrity keys, deploying older version, backdoor, etc. It suggests best practices such as code-reviews, automation, verifying artifacts, validating configuration, binary provenance, etc. The binary provenance verifies input to the artifact and validate transformation and entity that performed the build. The provenance fields include authenticity (signature), output, input (source and dependencies), command, environment, input metadata, debug-info, versioning. A build is considered verifiable if the binary provenance produced by the build is trustworthy. The verifiable build architectures include trusted build service, hermetic builds, reproducible builds, and verifiable builds, however you may need break-glass mechanism that bypasses the policy in case of outage. You can add post-deployment verification to validate the deployment.

The chapter fifteen shows how to investigate systems using debug flags, verifying data corruption, reviewing logs, and designing for safety.

The chapter sixteen is part of section four that focuses on disaster planning. This chapter introduces best practices to address short and long-term recovery such as performing analysis of potential disaster, establishing a response time, creating a response plans/playbooks, configuring systems, testing procedures/systems, and incorporating feedback from tests and evaluation. It shows how to setup incident response team that may include incident commander, SREs, customer support, legal, forensic, security/privacy engineers, etc. IR teams can use severity and priority models to categorize incidents based on severity of their impact on the organization and priority model to define response time. The response plan include incident reporting, triage, SLO, roles/responsibilities and communications. You also need to test systems and response plans and audit automated system. Red team testing can help simulate how the system reacts to an attack.

The chapter seventeen reviews crisis management that determines if the security incident is a crisis. This can be evaluated in triage that determines severity of the incident and whether the incident is a result of system bug or a compromise that is yet to be discovered. In the context of crisis management, operational security (OpSec) refers to the practice of keeping your response activity secret. For example, common OpSec mistakes include documenting incident in email, logging into the compromised systems, locking accounts/changing passwords, taking system offline. The chapter instead suggests meeting in person, use key-based access (without login), etc. You can apply forensics processes to investigate the security compromise. The chapter ends with summary of best practices that include triage, declaring an incident, communicate with executives and SecOps, creating IR team and forensics team, preparing communication and remediation and closure.

The chapter eighteen reviews recovery and aftermath from the security incident. You can establish recovery time based on if it affected mission critical system.

The goal of your recovery effort is to mitigate an attack and return your systems to their normal routine state, however complex security events may require parallelizing incident management/response execution. In order to return your systems to normal, you need to have a complete list of the systems, networks, and data affected by the attack. You also need sufficient information about the attacker’s tactics, techniques, and procedures (TTPs) to identify any related resources that may be impacted. There are several considerations before recovery such as:

  • how will your attacker respond to your recovery effort?
  • is your recovery infrastructure or tooling compromised?
  • what variants of the attack exist?
  • will your recovery reintroduce attack vectors?
  • what are your mitigation options?

The recovery checklists includes:

  • isolating Assets (quarantine)
  • system Rebuilds and software Upgrades
  • data sanitization
  • recovery data
  • credential rotation
  • postmortems

The chapter nineteen is part of section five of the book that offers making security a part of the organization culture. It suggests making security a team responsibility, providing security to users, designing for defense in depth and being transparent to the community.

The chapter twenty describes roles and responsibilities for security and reliability. For example, security experts implement security specific technologies, SREs develop centralized infrastructures, and security specialists can devise best practices. You can embed security experts with the development teams or review/audit security practices. Organizations can create red team that focus on offensive exercises for simulating attacks and blue team for assessing and hardening software and infrastructure.

The chapter twenty one shows how to build a culture of security and reliability. The chapter suggests organization culture of by-default security and reliability and encourage employees to discuss these topics early in project life-cycle. The chapter also suggests culture of review where peer reviews ensure that code implement least privilege and other security considerations. The culture should include awareness of security aspects, sustainability, transparency, and communication.

February 14, 2020

Review of the “Database Internals”

Filed under: Design,Technology — admin @ 8:06 pm

The database internals is an excellent resource for deep dive into storage engines and distributed systems. The first chapter introduces OLTP, OLAP and HTAP databases. It reviews database architecture and components including transport, query processor, storage engine, transaction manager, lock manager, access methods, buffer manager and recovery manager. The storage may use in-memory store or disk store and some in-memory database use disk for backup, which is updated asynchronously. The chapter reviews row-oriented and column-oriented databases along with data files and index files.

The second chapter covers B-Trees that is often used with disk based storage engines. The chapter introduces binary search trees (BST) and balance trees. However, such BST data structures use add elements in random order and are not optimized for disk storage as parent and child nodes can be stored in different regions of memory. Also, height of BST may limit the search in O(log N) operations. The chapter reviews architecture of hard drives such as SSD and B-Tree data structures where each node can hold up to N keys and N + 1 pointers to the child nodes. The nodes are grouped into root-node, leaf-nodes and internal nodes where each node is used for fixed-size page. Keys in B-Tree nodes are called index entries, separator keys or divider cells and they split the tree into subtrees holding key ranges. B-Trees are based on N logarithm base and there are K times more nodes on each new level. During lookup at most logk (M) (where M is total number of items) pages are fetched to find a searched key. In order to insert a value, it finds target leaf and key/value are appended to it. The node may need to be split if there isn’t enough room. Similarly, deletions find target leaf and key/value are removed. The deletion may result in node merges if neighboring nodes are too few.

The chapter three covers file format for B-Trees for disk. It reviews binary encoding and primitive types, strings and general principles of file format such as header, page-data and trailer. The page format can be fixed or variable size but variable size may incur more overhead. Also, variable size page must reclaim space when records are removed and reference records in page without regard to their exact locations. The variable-size pages generally use slotted page structure that has headers, list of pointers and list of variable size cells where each cell stores flags, key/data size, page-id and byte data. Removing an item may just mark the cell as deleted and reclaim later. The insertion may use first-fit or best-fit strategy to find the free blocks. Also, headers may store version and checksum for data validation.

The chapter four shows how to implement B-Trees, e.g. page-header may store flags, number of cells, magic number, etc. Some implementations of B-Trees store sibling pointers (forward/backward) to locate neighboring nodes but it adds complexity in split/merge. BTrees also store one additional pointer to child pages than the number of keys:

|                          |  Separator | ---> Ks >= K3
|  K1       |  K2            |    K3    |
Ks < K1    K1<=Ks<=K2     K2 <= Ks < K3

Alternatively, you can store rightmost pointer in the cell along with high key. Each node in B-Tree is designed to keep a specific number of items and resizing may require copying data so in order to avoid copying, they can use extension/overflow page and link it form the original page. B-Trees keeps keys in order so that they can use binary search and insertion point is index of the first element that is greater than the given key. Some implementations may store parent pointers in nodes or use breadcrumbs to store path of leaf node in case they need to split/merge. B-Tree implementations may postpone split/merge later, create a new right-most node or use other algorithms to improve re-balancing. B-Trees may also apply compression at various granularity levels and perform maintenance to fix fragmented data or garbage collect non-addressable data (vacuum).

The chapter five reviews transaction processing and introduces concepts of ACID and page caching so that modifications can be done in memory. The pages can be brought in if they are not in memory and evicted/flushed to disk when there isn’t enough memory (with O_DIRECT lag to bypass kernel cache). After page modifications, it’s marked as dirty so that it can be flushed for durability. These modifications are coordinated with the write-ahead-log (WAL) so that data can be recovered if the server crashes (referred as checkpoint). As splits/merge may require multiple writes, B-Tree  can lock pages that have high probability of being used, called pinning and pinned pages are kept in memory. The I/O operations can be buffered to reduce disk I/O. Based on available memory, B-Tree may need to evict old pages when new pages cannot fit in memory and there are a variety of algorithms for eviction policies (page replacement) such as FIFO, LRU, CLOCK (references in circular buffer), LFU, etc. B-Trees use write-ahead log (WAL) to buffer changes to page-contents. These changes to WAL are flushed with fsync, but due to certain error conditions in fsync it may not report errors if they were cleared and it can result in loss of data. B-Tree implementations may use seat/no-steal and force/no-force policies to determine when changes are flushed on disk and they impact undo/redo behavior. The steal policy allows flushing a page without committing a transaction and a force policy requires all pages modified by the transaction to be flushed before the transaction commits.  The chapter explains ARIES algorithm, which is steal/no-force recovery algorithm, uses physical redo to improve performance and logical undo to improve concurrency and uses WAL records to implement repeating history. ARIES uses LSN (log sequence numbers) to identify log records, track pages in dirty page table and use physical undo/logical undo. The chapter reviews concurrency controls such as optimistic concurrency control, multi-version concurrency control and pessimistic concurrency control (using lock and no lock). The chapter reviews transaction isolation and read/write anomalies such as dirty read (uncommitted updates), non-repeatable read (querying again), phantom read (range queries), lost update (last writer wins), dirty write (takes dirty reads), write skew (double spending). Th isolation include read-uncommitted that allows dirty, phantom and non-repeatable reads; read-committed that prevent dirty reads; repeatable that prevent non-repeatable reads but allow phantom reads; serializable level that executes transactions serially and prevent phantom reads. Serializable isolation is difficult to implement and some databases use snapshot isolation to observe all transaction committed since the start time. The snapshot isolation prevents lost update but it’s still susceptible to write skew. Optimistic concurrency control validates transaction before writing and works if retries can be prevented, but it still needs to manage a critical section. Multi-version concurrency control uses monotonically incremented transaction IDs or timestamps and is used to prevent access to uncommitted values. Pessimistic concurrency control can use locks or simple timestamps that it checks to ensure that no other transaction has been committed with higher timestamp. The database maintains max_read_timestamp/max_write_timestamp and read operations with older timestamp are aborted and write operations with lower than max_read_timestamp would conflict but write operations with older than max_write_timestamp are allowed (Thomas Write Rule). Lock-based concurrency control uses locks such as two-phase locking where growing phase all locks are acquired and shrinking phase, where all locks are released after the transaction. Locks can lead to deadlocks so you need timeout to abort long running transactions. The chapter describes distinction between locks and latches where locks are used to isolate and schedule overlapping transactions and latches guard physical B-tree contents (leaf/non-leaf). The latches can use reader-write locks (busy-wait/CAS) and latch crabbing determines to minimize holding time.

The chapter six goes over different types of B-Tree design and implementations. For example, some B-Trees use copy-on-write to copy contents in new shadow tree instead of using synchronization and latches and the pointer to top most page is atomically updated after the update (LMDB). In order to update the page on disk, the in memory representation is updated first using cached version, native pointers (unmanaged languages), language specific structures or using wrapper object to update disk as soon as B-Tree is updated. Lazy B-Trees reduce cost of updating, e.g. WiredTiger different format for in-memory and on-disk pages and updates are first saved in update buffer to reduce I/O. Lazy-Adaptive Tree group nodes into subtrees and attach buffer for batch operations to each subtree. FD-Trees append all changes to a small mutable head tree and multiple immutable sorted runs and use fractional cascading to maintain pointers between levels along with logarithmically sized sorted runs. In order to reduce write amplification, Buzzword-Tree (Bw) uses batch updates using append-only storage. Bw-Tree use compare-and-swap operations instead of synchronization. Cache-Oblivious B-Tree use cache-oblivious structures that give asymptotically optimal performance regardless of underlying memory structure. Cache-oblivious algorithms optimize two levels of hierarchy: page ache and disk and partition disk into blocks that page-location is cache aware. It uses platform parameters so that transfer between page-cache and disk is within constant factor.

The chapter seven discusses Log-Structured storage such as immutable LSM Trees that use append-only storage and merge trees. As B-Trees have high write amplification, LSM trees provide an alternative by using buffering and append-only storage. LSM Trees write immutable files and merge them together over time. LSM Trees use smaller in-memory buffer (memtable) and large disk. A separate write-ahead-log is appended and committed before in-memory operation is acknowledged to the client. After the disk flush, memory and disk sub-trees are discarded and replaced with the result of their merge. In LSM trees, redundant records are reconciled during the read and tombstones are used to mark deleted records. Some implementations use predicate deletes for range of keys to remove records. LSM may use compaction to optimize access such as leveled compaction used by RocksDB where level-0 tables are created by flushing memtable contents and then contents are merged later to create level-1. Some LSM trees use size-tiered compaction that group disk tables based on size or use time window for compaction (used by Cassandra). As opposed to B-Trees that are read-optimized, LSM trees do not require locating the record on disk during write but reads are more expensive with default configuration. The chapter then reviews sorted string tables (SSTables) that are often used to implement disk-resident tables. SSTables consists of index files and data files where index files use B-Trees or hash tables and data files holds data in key order and uses hash tables or other similar data structures for lookup/range queries. During compaction, data files can be read sequentially and merge iteration is order preserving so merge table can be created in a single run. The chapter introduces bloom filters test whether an element is a member of the set. The chapter then reviews Skiplist for keeping sorted data and use probabilistic balancing. A skip list builds hierarchy of linked-list at different heights where each node has more than one successor that point to nodes at lower-levels.

The chapter eight is part of second half of the book that focuses on distributed system. It introduces concepts of concurrency and parallelism where concurrent executions can interleave and shared state must be protected whereas parallel operations are executed by multiple processors. The chapter defines system reliability in terms of presence of fault tolerance and discusses fallacies of distributed computing (published by Peter Deutsch). In real applications, processing and latency time is not instantaneous and queue capacity is not infinite that also requires back-pressure. The queue size is determined by measuring task processing time and average time each task spends in the queue. Distributed system also have to deal with clock/time differences on multiple machines and state consistency such as read-time data repair or eventually consistent systems. Detecting failures in distributed systems is hard and requires heartbeat protocols and network partitions can result in partial failures. The chapter explains cascading failures that can propagate from one part of the system to another. You can use exponential backoff strategy and jitter to avoid amplifying problems. The messages can get lost, delayed or reordered in a distributed systems and sender may retry but it does not know if the message is already delivered, e.g. in fair-loss link a sender keeps retrying send infinitely; finite duplication won’t send messages finitely; and no-creation link won’t send the message the was never sent. Distributed systems use acknowledgments to notify the sender using sequence numbers and sender may re-transmit in absence of ack (stubborn link resend messages indefinitely). In order to prevent duplicate processing as a result of re-transmission, you can use idempotent operations. In distributed systems, messages can arrive out of order and recipient may use sequence to detect out of order message and put it in a buffer until earlier message arrives. The perfect link guarantees reliable delivery without duplication and no-creation (only deliver messages that were actually sent). Exactly-once delivery is very hard in distributed systems and most real applications use at-least-once delivery (at-most-once is not reliable). The chapter describes two-general’s problem to show link failures when communication is asynchronous even with perfect delivery as participants may not be alive or connected. This problem shows that no matter how many ACK you use, you can never be sure if message was delivered to both parties. This was further proved by FLP Impossibility problem that you can never guarantee consensus in a bounded time with asynchronous communication. The chapter finally discusses failure models such as crash faults, omission faults (skips execution of certain steps), arbitrary faults (byzantine faults), etc and you can use process groups and redundancy to mask these failures from user.

The chapter nine discusses failure detection, where a failure detector identifies failed or unreachable processes to exclude them  from the algorithm and guarantee liveness while preserving safety. Most distributed systems use heartbeats to detect failures, where the process notify its status to peers in response to heartbeat. Each process maintains a list of other processes and updates it with last response time. Some distributed systems use a deadline failure detector that uses heartbeat to detect if a process has failed to register within a fixed time interval. Alternatively, other systems use outsourced heartbeat to improve reliability using information from external perspective. Phi-Accural failure detector use phi-accrual failure detector to calculate probability of process’s crash based on sampling arrival time. Other approaches gossips by maintaining a heartbeat counter and sending heartbeat counter to random neighbor periodically. Another approach arranges active processes into groups where a process failure is detected by participants and the failure is propagated as a group failure.

The chapter ten goes over leader election while maintaining liveness, stability and safety. It starts with bully algorithm that uses process rank (e.g. biggest ip-address) to identify the new leader. However, it can be subjected to split brain and create problems if highest rank node is unstable. Next-in-line failover is another alternative where leader provides a list of failover nodes and next highest-ranked node is selected in case of leader failure. Candidate/Ordinary algorithm splits nodes into groups of candidates and ordinary, where one of the candidate node becomes a leader (picking highest-ranked alive node). Invitation algorithm allows processes to invite other processes to join their groups and smaller groups merged with bigger groups. Ring algorithm use ring topology where each process contacts its successor passing a set of nodes until one of the nodes respond. The highest-ranked node from live set is chosen as a leader. Lastly, you may use consensus algorithms to elect a leader along with failure detection algorithm.

The chapter eleven examines replication and consistency properties such as availability, fault tolerance and redundancy. The chapter reviews CAP theorem where availability requires non failing nodes to deliver results and linearizable consistency preserves the original operation order. In asynchronous system, you cannot guarantee both consistency and availability in presence of network partition so you either have to choose best effort availability or best effort consistency (or sacrifice latency). Also, Cap theorem discusses network partition where a node may serve incorrect response and not node crashes that doesn’t respond at all. The chapter reviews concepts of harvest and yield in context of CAP conjecture where harvest may return partial results and yield compares the number of requests succeeded  against the number of requests attempted. Thus, these properties focus on trade-offs as opposed to the absolute numbers. The distributed systems may abstract message passing and represent state as a shared memory where each unit of storage is called a register.  Each operation is tracked with invocation and completion event and the operation is considered failed if the process crashes before completing the operation. Also, some operations may overlap with other operations and are called concurrent operations. The registers can be categorized into safe (dirty/non-repeatable read), regular (repeatable), atomic (linearizable). The consistency model provide different semantics and guarantees from the perspective of state and operations in distributed/concurrent systems. For example, strict consistency provides complete replication transparency as if you hold a global lock but it’s impractical in real-life. Linarizability guarantees visibility of the writes to all readers exactly once without exposing partial state. If two operations overlap, all read operations occur after write operation can observe the effect of the operation. It provides total order of operations running concurrently so that every read of the shared value returns latest value written to the shared variable. The linearization point provides atomic guarantee such that the effect of operation becomes visible. Linearizability is expensive to implement as it requireds coordination and ordering but you can use compare-and-swap where you first prepare result and then use CAS for swapping pointers and publish the state. Sequential consistency is a step below Linearizability that executes operations in some sequential order where operations of each individual processes are executed in the same order. In causal consistency, all process see causally related operations in the same order. It can add logical clocks with each message and the operation is processed only if preceding operation is completed. The chapter defines vector clock as a structure for establishing a partial order between the events. Processes maintain vectors of logical clocks, with one clock per process and is incremented every time a new event arrives. In order to resolve conflict, you check duplicate value with same key and append a new version to the version vector and establish the causal relationships. The chapter discusses session models that evaluate consistency from the perspective of client and assume all client operations are sequential. It may use read-own-writes consistency model and monotonic read model that guarantees that you cannot read old value once you have seen new value. The monotonic write model guarantees that write of v2 follows write of v1. The write-follows-read ensures that writes ordered after writes that were observed by previous read operations. Eventual consistency propagates updates asynchronously and latest value is resolved using lat-write-wins or vector clocks. The eventually consistent systems provide parameters to tweak availability and consistency such as replication-factor (N), write-consistency (W) and read-consistency (R) and you can guarantee most recent value by using (R+W > N). You can optimize replication by grouping nodes into copy and witness subsets where witness replicas may store updates if copy replicas are running behind. The chapter ends with discussion of strong eventual consistency and CRDTs that are specialized data structures to guarantee consistency in any order. However, allowed operations have to be side-effect free, commutative, and causally ordered.

The chapter twelve discusses anti-entropy and dissemination of updates in context of broadcast, peer-to-peer and cooperative broadcast. The broadcast to all processes is expensive with large number of nodes and unreliable from a single process. The anti-entropy brings nodes back in sync in case of failures. Entropy measure disorder in the system and anti-entropy brings the nodes back up-to-date when delivery fails. The read repair detects and eliminate inconsistencies. It can be implemented as a blocking or asynchronous operation. Blocking read repairs ensures read monotonicity for quorum reads. Instead of issuing full read request from each node, the coordinator can issue one full read and send digest request to other replicas and then repair reads in case of inconsistencies. Another alternative is hinted-handoff, which is write-side repair mechanism where write coordinator stores hint record and replays to target node when it comes back. Some databases use sloppy quorum along with hinted-handoff where write operations use additional nodes that update crashed node when it comes back. Merkle Trees provide a compact hash representation of the local data. The replicas compare root-level hashes to check for inconsistency. Bitmap version vectors can also be used to resolve data conflict based on recency where logs of operations are kept on each node and are compared with other nodes and missing data is replicated to the target node. Gossip Dissemination use gossip protocols that are probabilistic communication procedure based on how rumors/diseases are spread. It use cooperative propagation to disseminate information where infective node spreads to susceptible nodes, which randomly update neighboring processes. Message redundancy metric is used to capture the overhead of repeated delivery and amount of time to reach convergence is called latency. Push/lazy-push multicast trees make a trade-off between epidemic and tree-based broadcast primitive by creating a spanning tree overlay of nodes to actively distribute messages with least overhead. It sends full message to a small subset of nodes and just message-id to rest and the node can query peer if it doesn’t have the data.

The chapter thirteen reviews distributed transactions. In order to make operations appear atomic, you may use atomic commitment algorithm that provides prepare, commit or rollback operations along with a transaction manager. For example, two-phase commit execute in two phases: prepare and commit/abort where a coordinator collects votes and rest of nodes called cohorts operate over disjoint datasets. In case of cohort failure, the coordinator will replicate decision values based on log. In case of coordinator failure, cohorts will not be able to learn the final decision. In order to make atomic commitment more robust against coordinator failure, three-phase commit adds extra step: propose, prepare and commit/abort. The transaction is aborted in case of coordinator failure or operation time-out. Next, the chapter reviews Calvin approach that uses deterministic transaction order to remove the need for coordination (as opposed to non-deterministic transaction in most databases that use two-phase or optimistic locking). For example, Calvin uses a sequencer that determines the order of transactions and establishes a global transaction input sequence and it may split time into epochs to minimize contention. The chapter discusses data partitioning and consistent hashing that map hashes to a ring and each node get its own position on the ring and becomes responsible for the range of values. If serializability is not required, you may use snapshot isolation that guarantees that all reads made within the same transaction are consistent with a snapshot of the database and only first committer wins when there is a write-write conflict. Lastly, the chapter discusses mechanisms to avoid coordination by preserving data integrity constraints.

The chapter fourteen discusses consensus that focus agreement, validity and termination (reach the decision). The chapter introduces concept of broadcast communication However, it may result in in-consistent state if the coordinator crashes while in the middle of broadcast. Atomic broadcast guarantee reliable delivery (atomicity) and total order. For example, virtual synchrony framework organizes processes into groups and messages to all its members are delivered in the same order. In Zookeeper atomic broadcast, a process takes the role of leader or follower and protocol splits timeline into epochs identified with monotonically increasing sequence number. The atomic broadcast is equivalent to consensus in asynchronous systems with crash failure. Paxos is commonly used algorithms that defines three roles: proposers, acceptors, and learners. It is split into two phases: voting (proposers compete for the leadership) and replication (proposer distributes values to acceptors). When acceptor receives prepare request, it can accept the proposal, respond with previously accepted message, notify proposer if local sequence number is higher. During replication phase, proposer can start the replication by sending Accept message to all acceptors. Paxos use quorum to make sure some participants can fail but still proceed as long as minimus number of votes required for the operation are available. Liveness is guaranteed in the presence of f failed processes and so that given 2f + 1 processes, f processes can fail and f + 1 processes can proceed. Multi-Paxos algorithm introduces role of a leader, a distinguished proposer to improve efficiency. The leader periodically contacts the participants to notify them it’s still alive with a lease timeout so that participants won’t select other leader until lease expires. Fast Paxos algorithm reduces a number of messages and let any proposer contact accepts directly rather than voting through the leader with total 3f + 1 processes. Egalitarian Paxos partitions the system into smaller segments and uses a leader for the commit of a specific command. Flexible Paxoes uses intersection of nodes that are used in propose and accept phase, .e.g given N participants, Q1 nodes for the propose phase to succeed and Q2 nodes for the accept phase to succeed, wen can ensure that Q1 + Q1 > N and Q2 can contain N/2 acceptors and Q1 = N – Q2 + 1. Next, the chapter discusses raft algorithm that makes concept of leader a first-class citizen that coordinates state machine manipulation and replication similar to atomic broadcast and Multi-Paxos that replicates multiple values instead of just one (a single leader makes atomic decisions and establishes message order). Each participant in Raft take the role of candidate, leader (for a term) and follower (similar to acceptor/learner). It divides time into terms/epochs to guarantee global partial order without relying on clock synchronization. Terms are monotonically increasing and each command is uniquely identified by the term number. During leader election, candidates send RequestVote message to other processes including candidate’s term and ID of the last log entry it observed. After collecting a majority of votes, the candidate is selected as the leader for the term. The Raft protocol uses periodic heartbeat to ensure the liveness of the participants and it may start new election after election timeout. The leader repeatedly append new values to the replicated log by sending AppendEntries message that include leader’s term, index and term of the log entry. A leader is elected only if it has the higher term ID than the follower. In case of split vote, Raft uses randomized timers to reduce the probability of multiple subsequent election ending up in a split vote. The leader sends heartbeat to the followers to detect failures and new election can be initiated if leader is down. The leader does not remove or reorder its log contents; it only appends new messages to it.

The chapter then reviews Byzantine consensus where distributed systems are deployed in adversarial environments that is prone to byzantine failures such as ill intentions, bugs, misconfiguration and data corruption. Most Byzantine consensus algorithms require N^2 messages to complete an algorithm step, where N is the size of the quorum. It discusses Practical Byzantine Fault Tolerance (PBFT) that assumes independent node failure but entire system cannot be taken over at once. All communication is encrypted and replicas know one another’s public keys to verify identities. PBFT guarantees both safety and liveness, no more than (n – 1) / 3 replicas can be faulty. For a system to sustain f compromised nodes, it is required to have at least n = 3f + 1 nodes. To distinguish between cluster configuration, PBFT uses view where in each view, one of the replica is a primary and the rest are backup. All nodes are numbered consecutively and the index of the primary node is v mod N where v is the view id and N is the number of nodes. The view can change when the primary fails. Clients execute their operations against the primary that broadcasts the request to the backup, which execute the request and send a response back to the client. The client waits for f + 1 replicas to respond with the same result for any operation to succeed. Replicas save accepted messages in a stable log and it is kept until it has been executed by at least f + 1 nodes. This log can be used for recovery in case of network partition but it is verified to prevent the attack vector. After every N requests, the primary makes a stable checkpoint, where it broadcasts the latest sequence number and waits for 2f+1 replicas to respond, which constitutes a proof for this checkpoint.

July 12, 2019

Review of “Designing Data Intensive Applications”

Filed under: Computing,Technology — admin @ 4:38 pm

Designing Data-Intensive Applications is one of best resource for building scalable systems. The book is full of comprehensive material on data processing and following is summary of essential lessons from the book:

The first four chapter covers foundation of data systems and the first chapter starts with basic terminology of reliability, scalability, maintainability, Operability, simplicity and evolvability. It recommends measuring latencies using percentile and defining SLO/SLA.

The second chapter reviews data model such as relational model, document/ graph model, etc. It reviews differences in query model SQL, graph queries, and SPARQL and compares them in terms of schema evolution and performance.

The third chapter introduces basic concepts of storage for relational and NoSQL databases. It shows how you can use hash indexes for looking up key-value data, which can be enhanced with compaction. It further reviews SSTable and LSM-tree structures. The Sorted-String Table or SSTable keeps key-value in sorted fashion where each key only occurs once within each merged segment file. These segments are later merged using merge sort and written to disk. The Log-Structured Merge-Tree (LSM tree) is used to merge and compact sorted files. In order to reduce lookup time, bloom filter can be used to ensure key exists in the database. The relational databases use B-Tree data structure that break the database into fixed size pages and each page is identified by address that are stored in a tree structure. The B-tree use write-ahead log (WAL) to persist new data before updating B-tree data structure. Another form of index is multi-column index that combines several fields into one key. The  chapter then differentiates between OLTP and OLTP and using star (fact table surrounded by its dimension tables) and snowflakes (dimensions are broken down into sub dimensions) schema for analytics. Column-oriented storage can also be used for data warehouse that leads to better compression, CPU cache usage and vectorized processing. Some data warehouse use materialized aggregates or data cube that provides aggregates grouped by different dimensions.

The fourth chapter covers encoding and serialization schemes such as JSON, XML and binary encoding including MessagePack and Avro. It reviews binary protocols of Thrift and Protocol Buffers. Further, it provides support of schema evolution in these protocols for backward and forward compatibility. It describes data flow via network exchange using SOAP, REST/RPC message brokers and distributed actor frameworks.

The fifth chapter is part of second part of the book that focuses on distributed data with emphasis on scaling and shared-nothing architecture. The fifth chapter describes replication and defines leaders and followers. It starts with basic leader-based replication such as active/passive or master/slave where writes go through leader and reads can be served by any node. The replication can be synchronous, asynchronous or semi-synchronous. Though, synchronous replication has performance issues but research in chain replication provides good performance, which is used by Azure storage. The chapter reviews scenarios of leader or follower node failure, which may require election of new leader but can be subjected to loss of data that wasn’t replicated from the old leader. The replication can use statement-based replication, shipping WAL logs, logical (row-based) replication or trigger-based replication, where former can have side-effects due to triggers/non-deterministic behavior. The version mismatch may cause incompatibilities with WAL based replication. Asynchronous cause effects of eventual consistency where latest reads are not fully replicated so you need read-after-write consistency that can be addressed using read-your-writes consistency, e.g. reading from the leader or remembering timestamp of last write. The lag in replication can display load data after showing new data and monotonic read consistency addresses that behavior. Consistent prefix read guarantee sequence of writes order is preserved when reading those writes. When replicating with multi-leaders across data-centers can have higher lag time between replication and may result in write conflicts between leaders. You may associate users to a specific leader based on location or resolve these conflicts using timestamps or higher unique id. Other forms of replication uses leaderless or quorum based consistency. Some implementation use asynchronous read-repair to fix stale data and anti-entropy process to add any missing data. The quorum based consistency uses odd number of nodes with w = r = ceil(n+1)/2, though it can also be subjected to stale values (with sloppy quorum and hinted handoff) or loss of data (last writer wins). The quorum based nodes may use version number for every key to preserve order of write (version vectors).

The chapter six covers partitioning works with replication so that each node might have more than one partition. Some partitions may be skewed having more data than others, also referred as hot spot. You can partition based on key-range or hash of key, where hashing may use consistent hashing to distribute keys fairly. The partitioning data also requires partitioning secondary indexes or maintaining a global index that covers all partitions by term. The partitioning may also require re-balancing where you may define 100 times more partition than nodes so that data is not only partially relocated. When reading data requires routing to nodes where client may contact any node directly, which forwards to other node if needed; send all requests to a routing node; or client is aware of partitioning and contacts appropriate node.

The chapter seven describes transactions and defines meaning of ACID, however it cautions against failures due to asynchronous writes, caching, disk failures (despite fsync), etc. The relational databases use transaction scope to write multiple objects atomically and in isolation and other databases use compare-and-set operations to prevent lost updates. Due to high cost of serializable transactions, most databases provide weak isolation level when running multiple transactions concurrently. In order to prevent concurrency bugs, database provide transaction isolation including read-committed for preventing dirty reads/writes that may use row-level locks and keeping copies of old data; snapshot isolation/repeatable read that prevents read skew (non repeatable reads) using reader/write locks and MVCC. Relational databases provide explicit locking using SELECT…FOR UPDATE to prevent lost updates, other databases use compare-and-set to avoid lost update. These transaction isolation can still lead to write skew and phantoms that can be prevented using SELECT FOR UPDATE, e.g. meeting reservation, choosing username, preventing double spending. The serializable isolation provides strongest guarantees but it’s not provided in most databases and has limitations with partitioned data. Two-phase locking can be used for serializable isolation but it suffers performance issues and can lead to deadlocks. Another alternative is predicate locks that locks all objects matching criteria but they also suffer from performance issues. Other alternatives include index-range locks and serializable snapshot isolation (SSI) that uses optimistic concurrency controls.

The chapter eight discusses network faults and failures in distributed systems. The chapter covers cloud computing, supercomputer where nodes communicate through RDMA/shared memory. These failures are common in most systems and can be detected by load balancer or monitoring system. Partial failures are hard to detect and you may use timeout to detect failures. Distributed systems may use monotonic clocks (System.nanoTime) or time-of-date (NTP) clocks, however unreliable clocks can make measuring time in distributed systems error prone. The NTP synchronization is not always reliable and drift in time-of-clock may result in incorrect order of events and last-write-win strategy may overwrite data with old value. Google TrueTime API uses confidence interval with clock time, e.g. Spanner uses clock confidence interval for snapshot isolation to create global monotonic increasing transaction ID. When scheduling task periodically, check process delay due to GC, virtualization, disk I/O, etc. When using a lock or lease, fencing ensure there is only one leader. When the lock server grants a lock/lease, it returns fencing token (monotonic number) and client includes it with each request so old requests are rejected. However, fencing token cannot prevent against Byzantine faults (deliberate faults).

The chapter nine explains consistency and consensus. Most replicated database provides eventual consistency, which is weak consistency. Linearizability is strongest consistency that makes system appear as if there is a single copy of the data and reads cannot return old data if it previously returned new data. Linearizability may use compare-and-set operation to prevent data overwrite. The chapter also distinguishes Linearizability with Serializability that is isolation property of transaction that guarantees serial order of transactions, where Linearizability guarantees recent data after read/write and doesn’t prevent write skew. Leader election, distributed locks, unique constraints such as username may use Linearizability to come up with a single up-to-date value. Simplest way to have linearizable systems is to keep a single copy of the data but you need replication for fault tolerance system. You can use single-leader replication where all writes go to leader or consensus algorithm but multi-leader and leaderless replication (dynamo style) don’t provide linearizable guarantees. The chapter then describes CAP theorem, where consistency (C) relates to linearizability and you give up availability if some replicas are disconnected and wait until replicas are fixed. On the other hand, a replica can remain available even if it’s disconnected from other replicas to provide availability at the cost of linearizability. Also, CAP theorem only considers one kind of fault – network partition or nodes that are alive but disconnected from each other so most highly available systems don’t meet CAP definition. Though, causal order can define what happened before what but it’s not total order guaranteed by linearizability and sets can only be partially ordered. Linearizability is stronger than causality but most systems only need causality consistency that show what operation happened before other operation. Sequence number or timestamp (logical clock) can generate sequence number to identify order of operations, which can be incremented by the single leader. Other systems can preallocate blocks of sequence, attach timestamp or generate local sequence number but they are not consistent with causality. The chapter then reviews Lamport timestamps (logical) that enforce order (distinct from version vectors). You need a leader to sequence all operations on a single CPU to guarantee total order broadcast (atomic), but it’s not scalable. A leader per partition can maintain ordering per partition but it doesn’t guarantee total ordering. Total order broadcast requires reliable delivery and totally ordered order (same order to all nodes) and it can be used to implement serializable transaction. It can also be used for implementing lock service that provides fencing tokens. You can implement this by appending message to log, reading the log and waiting for the message delivered (same order for all nodes) back to verify the unique identifier such as username. But it only guarantees linearizable writes and you must sequence reads only after message is delivered back to you to guarantee linearizable reads.

Next part of chapter nine describes consensus, where you have to agree on leader after election (in presence of network faults while avoiding split brain) and atomic commits (as in ACID). Despite FLP that proved no algorithm can reach consensus if a node can crash, distributed systems can achieve consensus. The chapter reviews Two-phase commit (2PC) that involves multiple nodes as opposed to a single node that uses logs the data before committing it. The 2PC uses coordinator/transaction manager that creates globally unique tx-id and tracks two phases: prepare and commit. The coordinator must write transactions in logs before commit in case the coordinator crashes while nodes may have to wait for coordinator indefinitely. Three-phase commit assumes bounded delay/timeout to prevent blocking atomic commits. The chapter then distinguishes between internal database and distributed transactions, where distributed transactions guarantee exactly-once processing atomically such as XA transactions. On the downside, coordinator in distributed transactions would be a single point of failure and they limit scalability. The fault tolerant consensus requires uniform agreement, integrity, validity, and termination to agree on same value. The best algorithm that provides fault tolerant consensus include VSR, Paxos, Raft and Zab that uses total order broadcast algorithms that requires messages be delivered exactly once in the same order to all nodes. These algorithms use epoch numbers (monotonically increasing) for each election and quorum is used to agree on the value. There are a few limitations of consensus such as synchronous replication, majority voting (minimum 3 nodes), static membership, continuous re-election due to partial failures. Some of implementations include Zookeeper and etc that provides linearizable atomic operations using compare and set to implement locks; total ordering of operations using fencing token (monotonically increasing); failure detection; change modifications.

The chapter ten describes batch processing and is part of third part of the book that covers data derived from system of record for OLAP and reporting. The batch processing is offline processing as opposed to real-time services or near-real-time stream processing. The chapter starts with basic Unix tools that uses pipes and files for batch processing. It then reviews MapReduce such as Hadoop and distributed file systems such as HDFS, GlusterFS, QFS. The mapper extracts key/value from input records and generate a number of key-value pairs, whereas reducer takes key-value pairs and collects all values belonging to the same key. The scheduler uses principle of “putting the computation near the data” to run mapper on the machines with replica of input file. MapReduce often requires workflow systems to manage dependencies such as Oozie, Azkaban, Luigi, Airflow, Pinball. These frameworks use groups to merge or collate related data. However, you can use skewed join when group data is very large to fit on a single machine. where work can be parallelized on multiple reducers. Mappers can use broadcast hash joins and partitioned joins when working with large data. The chapter also discusses data flow engines like Spark, Tea, Flink that supports operators for providing more flexible way to create data pipeline. It also reviews graph processing systems such as Apache Graph, Spark GraphX that supports bulk synchronous parallel (BSP) model of computation, which sends messages from one vertex to all connecting vertices similar to actor model if you think of each vertex as actor.

The chapter eleven discusses stream processing for providing near real-time processing. The stream processing transmit event streams using messaging systems and pub/sub model. The messaging system may use UDP, brokerless libraries such as ZeroMQ or message brokers with queue support. The message brokers support load balancing when a single message is delivered to one of consumer and share the work (shared subscription). Alternatively, they use fan-out where message is delivered to all consumers. Message brokers use ack to remove the message when it’s processed by consumer. The broker delivers the message again in event of connection failure using atomic commit protocol. Some message brokers such as Apache Kafka use append-only logs to add incoming messages and these logs can be partitioned where each message uses monotonically increasing sequence number (without ordering gurantee). The consumers read files sequentially by specifying offset. There are some limitations of these log based brokers, e.g. number of nodes sharing the work can be at most the number of log partition in that topic and if a single message is slow to process, it holds up processing in that partition. Some implementations may use circular buffer to store messages on disk but if consumers cannot keep with producers, it may drop old messages. The chapter discusses change data capture for syncing data that capture the changes in the database and apply same changes to search index. You can also use compact logs for syncing the data from offset 0 and scan over all messages. The chapter reviews event sourcing that stores all changes to the application state as a log of change events. The event source distinguishes between event and commands and after command is validated, it becomes an event that is durable and immutable. The immutable data is related to command query responsibility serration principle.

The last chapter reviews future of data systems and limits of total order that may require a single leader but it can be difficult in distributed data-center and micro services. You can send all updates to the same partition but that is not sufficient to capture causal dependency. The chapter covers lambda architecture and unbundling of databases including storage technologies such as secondary indexes, replication logs, text search indexes. It reviews exactly-once execution of operation, duplicate suppression and operation identifier (using request-id to make operation idempotent).

June 18, 2019

Review of “Designing Distributed Systems”

Filed under: Computing,Technology — admin @ 11:34 am

The “Designing Distributed Systems” book provides design patterns for building distributed systems with support of container technologies such as Kubernetes. The book consists of three sections where first section focuses on single-nodes, second section focuses on long-running services, and third section focuses on batch computation.


The first pattern in the book introduces concept of sidecar pattern for modularity and reusability where a single application requires two containers: application container and sidebar container where sidebar container provides additional functionality such as adding SSL proxy for the service, collecting metrics for the application container. The side-bar container can be configured via dynamic configuration service.


The ambassador pattern introduces an ambassador container that sits between the application and external services and all incoming/outgoing traffic goes through it. It also helps with modularity and reusability where the ambassador may abstract sharded service (or A/B testing) so that client or service itself doesn’t need to know all details . You may also use ambassador container for service brokering where it looks up an external service and connects to it.


The adapter pattern uses special container to modify the interface of application container, e.g. you can deploy monitoring adapter to automatically collect health metrics using Prometheus or other tools. Similarly, you may use adapter container to collect kubernetes logs (stdout/stderr) and reformat the logs before sending them to log aggregator (Fluentd).

Replicated Load-Balanced Services

This pattern is part of long-running services where a load balancer is added in front of the service for scalability. Each service is designed as a stateless so that requests can be sent to any replica of the service behind the load balancer. Each service needs to provide readiness probe so that load balancer knows if it can serve the requests. In some cases, you may need to support session-tracked services where user requests are routed to the same replica using sticky session or consistent hashing function. You may add a caching layer that is deployed along with your service container (as sidebar). Further, you may need to provide rate-limiting and protect against DOS attacks (X-RateLimit-Remaining headers). This pattern can also implement SSL Termination where external traffic is encrypted with different certificate compared with internal traffic (Varnish).

Sharded Services

This pattern partitions the traffic where each shard serves subset of all requests. As opposed to replicated services that are generally used for stateless services, sharded services are used for building stateful services. You may use sharded cache for each shard that sits between user and front-end to optimize end-user performance and latency. You may add replicas for each shard for further redundancy and scalability. Sharding requires selecting a key to route the traffic, e.g. you may use IP-address or consistent hash function to avoid remapping when new shards are added. If one of the shard becomes hot, you can add replicated sharded cache to handle the increased load.


The scatter/gather pattern adds parallelism in servicing requests where work is broken and spawned to multiple services and then result is aggregated before returning to the user. For example, you can implement distributed document search by farming multiple leaf machines that returns matching document and root node aggregates the results. You can also add support for sharded data by searching each shard in parallel and root node generates union of all documents returned by each shard (leaf node). One downside of this pattern is that it may suffer straggler problem as total response time depends on the slowest response so you may need to replicate each shard to improve computational power.

Functions and Event-Driven Processing

This pattern is used to implement function-as-a-service (FaaS) products. FaaS simplifies development and deployment as the code is managed and scaled automatically. However, FaaS requires that you decouple your application into small parts that can be run independently. Faas uses event systems to communicate with each function or create a data pipeline. You can use external data services for storing states that is shared by these functions.

Ownership Election

This pattern helps in multi-node environment where a specific task must be owned by a single process. For example, when you have multiple replicas, you may need to elect master using consensus algorithm such as Paxos, Raft or frameworks such as etcd, ZooKeeper, and consul. You can use distributed locks to implement ownership (optionally with a lease or TTL). You may need to verify if you hold the lock before proceeding, e.g.

func (Lock l) isLocked() boolean {
return l.locked && l.lockTime + 0.75 * l.ttl > now()

Work Queue Systems

This pattern is part of batch computation section to handle work items within a certain amount of time. You may use a work-queue manager container along with an ambassador container to connect to external queue source where source might use storage API, network storage, pub/sub systems like Kafka or Redis. Once the queue manager receives a work item, it launches a worker container. Kubernetes contains a Job object that allows for the reliable execution of the work queue. In order to limit number of worker containers running concurrently, you can limit the number of Job objects that your work queue is willing to create. You can also use the multi-worker pattern when different worker containers are transformed into a single unified container that implements the worker interface.

Event-Driven Batch Processing

This pattern allows data pipelining where an output of one work queue becomes input to another work queue, referred as workflow systems. Here are patterns of event-driven processing:
Copier: This pattern just duplicates the work item into two or more identical streams.
Filter: This pattern reduce a stream of work items to a smaller stream of work items by filtering out items that don’t meet particular criteria.
Splitter: This works like filter, but instead of eliminating input, it sends different inputs to different queues based on criteria.
Sharder: This is more generic form of splitter and splits a work item into smaller work items based on sharding function.
Merger: This is opposite of copier and merges two different work queues into a single work queue.

You may use pub/sub API to communicate between different workers.

Coordinated Batch Processing

This is similar to Reduce part of MapReduce pattern where a work is broken up and distributed to multiple nodes in parallel. You may need Join or Barrier Synchronization to wait for intermediate results before proceeding to the next stage of the workflow. For example, reduce phase aggregates merges several outputs into a single output.

March 6, 2018

Tips from the second edition of “Release It!”

Filed under: Design,Methodologies,Technology — admin @ 4:19 pm

The first edition of “Release It!” has been one of most influential books that I have read and it introduced a number of methods for writing fault tolerant systems such as circuit-breaker, bulkhead patterns so I was excited to read the second edition of the book when it came out. Here are a few tips from the book:

In the first chapter, the author defines stability in terms of robustness, i.e., “A robust system keeps processing transactions, even when transient impulses, persistent stresses, or component failures disrupt normal processing.” He recommends focusing on longevity bugs such as resource leaking, e.g. you can set timeout when invoking a network or database operation or detect dead connections before reading/writing. The author cautions against tightly coupled systems that can easily propagate failures to other parts of the system. Similarly, high number of integration points can increase probability of failure in one of those dependencies. The author suggests looking into the blocking calls that can lead to deadlocks when using multiple threads or scrutinizing resource pool that gets exhausted from blocking operations. Another cause of instability is chain reaction from failure of one of servers that increases load on the remaining servers, which can be remedied using bulkhead or circuit-breaker patterns. High memory usage on the server can also constrain resources and the author recommends using external caching systems such as Redis, memcache, etc. In order to monitor health of the system, the author suggests using a mock transaction to ensure it’s working as expected and keeping metrics on errors especially login errors, high latency warnings. One common self-inflicting failure can be caused by self-denial attacks by marketing campaign that can be mitigated by using shared-nothing architecture or reducing fan-in shared resources. The servers can also be subjected to dogpile effect, where resources after upgrade, cronjob or config change spike up that can be mitigated by using random clock skew, adding random jitter and using exponential backoff. Finally, the author recommends monitoring slow responses, removing unbounded result sets and failing fast.

Here is a summary of stability patterns that the author recommends:

  • Apply timeout with integration points and delayed retries when failure occurs.
  • Apply circuit-breaker to prevent cascading failures (along with timeout).
  • Apply bulk-head pattern to partition the system in the event of chain reaction failure.
  • Steady state
  • Data purging
  • Log files
  • Fail fast, restart fast and reintegrate
  • Let it crash with limited granularity, e.g. boundary of actor.
  • Supervision for monitoring, restarts, etc.
  • Shed Load
  • Back-pressure – queues must be finite for finite response time
  • Governor – create governor to slow rate of actions (from automated response) so that humans can review it.

Next few chapters focus on network, machines and processes for building and deploying the code. The author offers several mechanisms for scaling such as load balancing with DNS, using service registry for upgrading and fail-over, configuration, transparency, collecting logs and metrics, etc. The author recommends load shedding when under high load or use HTTP 503 to notify load balancer. For example, queue length can be calculated as: (max-wait-time / mean-processing-time + 1) * processing-threads * 1.5. You can use listen reject queue to return 503 error to prevent clients from reconnecting immediately. The control-panel chapter recommends tools for administration. It recommends postmortem template such as what happened, apologize, commit to improvement and emphasizes system failures (as opposed to human errors). The author recommends adding indicators such as traffic indicators, business transaction, users, resource pool health, database connection health, data consumption, integration point health, cache health, etc.

The security chapter offers standard best practices from OWASP such as using parameterized queries to protect against SQL injection, using high entropy random session-ids/storing cookies for exchanging session-ids to protect against session hijacking/fixation. In order to protect against XSS (when user’s input is rendered in HTML without escaping) by filtering input and escaping it when rendering. The author recommends using a random nonce and strict SameSite policy to protect against CSRF. Similarly, author recommends using the principle of least privilege, access control, etc. The admin tools can offer tools for resetting circuit breakers, adjust connection pool sizes, disabling specific outbound integrations, reloading configuration, stopping accepting load, toggling feature flags.

For ease of deployment, the author recommends automation, immutable infrastructure, continuous deployment, and rolling changes incrementally.

The author suggests several recommendations on process and organization such as OODA loop for fast learning, functional/autonomous teams, evolutionary architecture, asynchrony patterns, loose clustering and creating options for future.

Lastly, the author offers chaos engineering as a way to test resilience of your system using Simian army or writing your own chaos monkey. In the end, the new edition offers a few additional chapters on scaling, deployment, security, and chaos engineering and more war stories from author’s consulting work.

January 15, 2017

Review of “Whiplash: How to Survive Our Faster Future”

Filed under: Future,Technology — admin @ 4:38 pm

I read “Whiplash: How to Survive Our Faster Future” by Jai Ito and Jeff Howe over the holidays. Joi Ito is a director of the MIT Media Lab. The MIT Media Lab was created by Nicholas Negroponte in 1985 to build an environment where best ideas from schools of arts and science can be married to build next revolutionary discoveries.

This book narrates anecdotes of how technology revolutionized human development in past and how it continues to disrupt our lives today. As a consequence of Moor’s law and the Internet, technology is changing at an exponential speed. In such rapidly changing environments, this book provides key lessons that can be used to prepare us for uncertain future and paradigm shifts. In such exponential times, the invention of new technologies far outpaces the moral and ethical consequences of those breakthroughs. As technologies can be used for both good and bad, they offer both salvage and demise of humanity.

Here are primary principles that authors present to shape the new world:

Emergence over Authority:

The invention of Internet has facilitated communication and collaboration all over the world, where best ideas can be easily shared and exchanged. As a result, institutes that had central authorities are disintegrating. The authors present several examples of Emergence vs Authority such as Blogs vs Newspapers, Wikipedia vs Encyclopedia and central governments vs social networks based political revolutions.
In emergent systems, participants use simple rules and exchange information to build complex systems. Examples of emergent systems are ant colony, slime mold, brain, flocking birds, stock exchanges, biology, etc. The authoritarian systems enable incremental changes whereas emergent systems are more adoptive and foster non-linear progress.

Pull Over Push:

Push-based systems control their access whereas pull based use transparency and two-way communication and are able to cope with the crisis far better than push based systems. The authors recited an example of pull-based when meltdown of Fukushima nuclear plant occurred as a result of severe earthquake and fourteen feet tsunami. Joi and a team of volunteers across the world collaborate and built Geiger counters to take accurate readings of radiation. Other examples of pull based systems are crowdfunding and crowdsourcing.

Compasses over Maps:

A map has detailed knowledge and an optimal route whereas compass offers more autonomy and offer more flexibility in an unpredictable environment. The authors stated examples of the education system where standardized tests and curriculum deprive students of creativity and passion for learning. They used the culture of Media Lab as an illustration where the vision is based on compass heading. It provides a framework for individual progress leaving flexibility for interactions between groups.

Risk over Safety:

Traditional businesses are more risk averse where new ventures are thoroughly analyzed and they spend millions in studies. However, the cost of experimenting new ideas has drastically been reduced in today’s market and it offers much better return on investment.

Disobedience over Compliance:

Innovation requires creativity and breaking rules so a high-impact institutes require a culture of disobedience. It needs a culture where criticism and diverse ideas are embraced.

Practice over Theory:

Due to low-cost of launching new products, an innovating organization requires a culture where experiments are valued more than detailed planning.

Diversity over Ability:

The authors provided lessons from biochemistry companies that used gamers to design protein molecules. They used gamers with diverse background and those gamers had better pattern recognition than the biochemists with PhD. Most organizations believe in diversity, but most organizations lack diversity especially in high-tech companies such as Facebook, Yahoo and Google.

Resilience over Strength:

Resilient organizations are like the immune system that can successfully recover from failures. The authors gave examples of cyber-security where there are threats from various sources and successful defense requires treating security systems as biological systems and building strong immune systems against those security risks.

Systems over Objects:

Systems over objects emphasize understanding the connections between people, communities, and the environment. Instead of optimizing an individual or an organization, we need to optimize the impact of innovations on an entire natural system.


In this final chapter, authors gave examples AI and machine learning where deep learning and reinforced learning has allowed machines to beat human experts in Chess and Go. The authors cite “The Singularity is Near” by Ray Kuzweil, who predicts that we will have intelligent explosion by 2045. In this world, we will have to think about how humans and machines will work together.

Powered by WordPress