A vocabulary,
in one place.
Across all the volumes — guides, simulators, papers, foundations, data structures — the same terms keep showing up. This is the index of them. One definition each, a longer explanation, and a link back to where the term lives.
- ACID Storage
Atomicity, Consistency, Isolation, Durability — the database transaction guarantees.
Coined by Andreas Reuter and Theo Härder in 1983. Atomic = all-or-nothing; Consistent = invariants hold; Isolated = concurrent txns appear serial; Durable = once committed, survives crashes.
See: Guide › transactionsRead full entry- AEAD · Authenticated Encryption with Associated Data Security
Encryption that integrates confidentiality and integrity in one operation.
AES-GCM and ChaCha20-Poly1305 are the canonical examples. Replaces the error-prone "encrypt-then-MAC" composition that gave us Lucky13, BEAST, and POODLE. Every TLS 1.3 cipher suite is AEAD.
See: Guide › httpsRead full entry- ALPN · Application-Layer Protocol Negotiation Networking
TLS extension that picks h2 / h3 / http/1.1 inside the handshake.
Avoids a second round trip after TLS finishes. Also how WebRTC and gRPC tell the server what protocol to switch to.
See: Guide › httpsRead full entry- Amdahl's Law Performance
Speedup = 1 / (s + p/N), where s is the serial fraction.
A 5% serial fraction caps speedup at 20× no matter how many cores you add. Why "make it parallel" doesn't scale forever.
See: Foundations › foundationsRead full entry- Anti-entropy Distributed
Background process that reconciles divergent replicas.
Merkle-tree comparison or read-repair. Cassandra and Riak run anti-entropy continuously to converge eventually-consistent replicas.
Read full entry- Anycast Networking
One IP, advertised from many locations; routing picks the nearest one.
How DNS root servers work, how Cloudflare and AWS Global Accelerator route traffic, how DDoS mitigation absorbs traffic across regions.
See: Guide › dns· Guide › cdnRead full entry- API Gateway Architecture
A traffic entry point that routes, authenticates, and rate-limits.
AWS API Gateway, Kong, Tyk, Envoy. Sits in front of microservices to provide a single client-facing surface.
See: Guide › api gatewayRead full entry
- B-tree Storage
Balanced multi-way search tree; the shape of every relational index.
Bayer & McCreight, Boeing 1971. Wide fan-out keeps the tree shallow even at billions of records — 4 levels deep at fanout 200 vs 30 levels in a binary tree.
See: Simulators › btree· DS › data structuresRead full entry- Backoff Architecture
Wait longer between retries after each failure.
Exponential, jittered, decorrelated. Without jitter, all clients retry at exactly the same moment — the thundering herd that crashes a recovering service.
See: Simulators › retry strategyRead full entry- Backpressure Architecture
A signal from a slow consumer asking the producer to slow down.
Without it, queues grow until OOM. Reactive Streams and HTTP/2 flow control bake it in; most systems retrofit it badly.
Read full entry- Bayes' theorem ML
P(A∣B) = P(B∣A)·P(A)/P(B). The rule for updating belief.
Counterintuitive when P(A) is small: a 99%-accurate test on a 1-in-1000 disease still gives ~9% true-positive rate.
See: Foundations › foundationsRead full entry- BBR · Bottleneck Bandwidth and RTT Networking
Congestion control that models the path instead of reacting to loss.
Cardwell et al at Google, 2017. Replaces CUBIC's "halve cwnd on loss" with active probing of bandwidth and RTT. Default in Linux 4.9+ and Google's edge.
See: Papers › papersRead full entry- BFF · Backend-for-Frontend Architecture
A per-client backend that aggregates and shapes services for one frontend.
A web BFF, a mobile BFF, a partner BFF. Keeps OAuth refresh tokens out of browsers (the Part 10 OAuth pattern). Adds a hop; saves a thousand frontend headaches.
See: Guide › oauthRead full entry- BGP · Border Gateway Protocol Networking
How autonomous systems on the internet exchange routes.
Path-vector protocol with policy. Most internet outages that "took down half the internet" were BGP misconfigurations.
See: Guide › bgpRead full entry- Birthday bound Security
Hash collisions become likely at √(output space).
A 256-bit hash gives only ~128 bits of collision-resistance. Why MAC tags need to be at least 128 bits in modern protocols.
See: Foundations › foundationsRead full entry- Bloom filter Storage
Probabilistic set: "definitely not" or "probably yes" in tiny space.
Burton Bloom, 1970. Cassandra row caches, every LSM-tree, Bitcoin SPV, Chrome safe-browsing. ~10 bits per element for 1% false-positive.
See: Simulators › bloom filter· DS › data structuresRead full entry
- Cache hit ratio Performance
Fraction of reads served from cache.
95%+ is a healthy target for read-heavy services. Below 80% suggests cache key design is wrong.
See: Handbook › caching strategiesRead full entry- Cache stampede Performance
When a popular cache key expires and many concurrent requests miss simultaneously.
Each miss does the expensive backend work; backend gets crushed. Mitigations: probabilistic early expiration, request coalescing, lock-on-miss.
See: Handbook › advanced cachingRead full entry- Cache-aside Architecture
App reads cache; on miss reads source and populates.
The most common caching pattern. Simple to reason about; staleness is the cost.
See: Handbook › caching strategiesRead full entry- CAP theorem Distributed
Under partition, choose Consistency or Availability.
Eric Brewer, 2000. Often misread — it's about behaviour during a partition, not a permanent trade-off. Most real systems are AP with bounded staleness.
See: Simulators › cap theoremRead full entry- Cardinality Storage
Number of distinct values in a set.
High-cardinality columns make poor partition keys (uneven distribution); low-cardinality columns make poor indexes (each key matches many rows).
See: Foundations › foundationsRead full entry- CDN · Content Delivery Network Networking
A global cache fabric with PoPs near users.
Cloudflare, Akamai, Fastly. Modern CDNs do more than caching — TLS termination, WAF, edge compute, image transforms.
See: Guide › cdnRead full entry- Certificate Transparency · CT Security
Append-only public logs of every TLS certificate issued.
RFC 6962. Browsers refuse certs without a Signed Certificate Timestamp embedded. Tools like crt.sh let you spot rogue certs for your domain in minutes.
See: Guide › httpsRead full entry- Cgroup Operations
Linux mechanism that limits and accounts CPU, memory, IO for a process tree.
Plus namespaces = container. cgroupv2 unified the interface in 2016; required for systemd and modern container runtimes.
Read full entry- Chain replication Distributed
Replicas in a linear chain; writes flow head→tail, reads from tail.
van Renesse & Schneider, 2004. Strong consistency with simpler failure handling than Paxos. Used in Microsoft's Azure storage and FoundationDB.
Read full entry- Choreography vs orchestration Architecture
In a saga: services react to each other's events vs a central coordinator drives steps.
Choreography is decentralised, harder to debug. Orchestration centralises the workflow; easier to reason about, single point of failure.
Read full entry- CIDR · Classless Inter-Domain Routing Networking
IP address + prefix length notation (e.g. 10.0.0.0/24).
Replaced the rigid Class A/B/C system in 1993. The /N tells you how many high bits identify the network.
See: Tools › cidr calculatorRead full entry- Circuit breaker Architecture
A state machine that fails fast after a downstream is unhealthy.
Closed → Open → Half-open. Stops you from beating up an already-broken service. Hystrix popularised it; Resilience4j is the modern Java incarnation.
See: Simulators › circuit breakerRead full entry- CLOCK Algorithms
Approximate-LRU cache eviction with a circular reference bit.
Each entry has a referenced bit. Eviction sweeps a clock hand; sets the bit to 0; evicts when it sees a 0. Used by Redis (allkeys-lru), Postgres buffer pool.
See: Simulators › cache evictionRead full entry- Compaction Storage
Merging LSM SSTables to bound read amplification.
Tiered (size-tiered): merge files of similar size. Levelled (RocksDB default): each level is K× the size of the previous, single sorted run per level.
See: Guide › kafkaRead full entry- Connection pooling Architecture
Reuse a fixed pool of long-lived connections instead of opening one per request.
Database connections are expensive; PgBouncer, HikariCP, and the like multiplex thousands of short requests over hundreds of long connections.
Read full entry- Consistent hashing Distributed
Hash both keys and nodes onto a ring; nodes own arcs.
Karger et al, MIT 1997. Adding/removing a node only moves K/N keys, not the whole keyspace. Powers every distributed cache and CDN.
See: Simulators › consistent hashingRead full entry- CQRS · Command Query Responsibility Segregation Architecture
Separate read and write models, often backed by different stores.
Pairs naturally with event sourcing. Read side can be denormalised, indexed, or materialised differently from the write side.
Read full entry- CRDT · Conflict-free Replicated Data Type Distributed
A data type whose merges are commutative, associative, idempotent.
Strongly eventual consistency without coordination. The math behind every collaborative editor (Yjs, Automerge).
See: Simulators › crdt· Papers › papersRead full entry- CSP · Content Security Policy Security
HTTP header that whitelists which origins can load scripts, styles, frames.
The single most effective XSS mitigation when configured strictly. nonce or hash-based for inline scripts.
Read full entry- CSRF · Cross-Site Request Forgery Security
Attack: a third-party site causes the browser to send authenticated requests to your origin.
Defenses: SameSite cookies, anti-CSRF tokens, double-submit cookies. Largely solved by SameSite=Lax becoming the browser default.
Read full entry
- DAG · Directed Acyclic Graph Algorithms
Edges have direction; no cycles.
The natural shape for build systems, version control, schedulers, dataflow engines. Topological sort produces a valid execution order in O(V+E).
See: DS › data structuresRead full entry- DAX Storage
AWS DynamoDB Accelerator — an in-memory cache in front of DynamoDB.
Sub-millisecond reads. Pattern transferable: every fast KV store benefits from a memory tier.
Read full entry- DDoS · Distributed Denial of Service Security
Coordinated traffic flood from many sources.
L3/L4 (volumetric, SYN floods) vs L7 (HTTP slow-loris, regex). Anycast + scrubbing centers (Cloudflare, AWS Shield) absorb the volumetric kind.
Read full entry- Deadlock Concurrency
Two or more transactions wait on each other forever.
Detected by cycle-finding in a wait-for graph; resolved by aborting the youngest transaction. Postgres detects deadlocks every 1 second by default.
Read full entry- Distributed tracing Operations
A request's path across services, captured as a tree of spans.
OpenTelemetry standardised the wire format. Sampling is the open problem at scale — head-based vs tail-based.
Read full entry- DPDK Networking
Data Plane Development Kit — userspace networking that bypasses the kernel.
Polled-mode drivers. Hundreds of millions of packets per second on commodity hardware. Used by load balancers, telecom NFV, high-frequency trading.
Read full entry- DPoP · Demonstrating Proof of Possession Security
Bind an OAuth token to a client-side public key.
RFC 9449. Theft of just the access token is useless without the matching private key. Replacing bearer tokens for high-value APIs.
See: Guide › oauthRead full entry
- eBPF Operations
Run sandboxed programs in the Linux kernel without loading modules.
Originally for packet filtering (Berkeley Packet Filter); now used for tracing, networking (Cilium), security (Tetragon), observability.
Read full entry- Edge compute Architecture
Run code at the CDN edge, milliseconds from the user.
Cloudflare Workers, AWS Lambda@Edge, Vercel Edge Functions. V8 isolates rather than full processes — cold start measured in microseconds.
See: Guide › cdnRead full entry- Epoll Concurrency
Linux IO multiplexing — register fds, get notifications on readiness.
Replaces select/poll which were O(N) per call. epoll_wait is O(1). The foundation of every async Linux network server.
Read full entry- Erasure coding Storage
Store k data shards + m parity shards; tolerate any m losses.
Reed-Solomon for cold storage. ~1.5× space overhead vs 3× for replication, at the cost of more compute on read with failures.
Read full entry- Event sourcing Architecture
Persist state as an append-only log of events; current state is a fold.
Audit, replay, and time-travel debugging come for free. The cost is querying — you usually pair with CQRS read models.
Read full entry- Eventual consistency Distributed
All replicas eventually agree if writes stop.
The default for AP systems under CAP. "Eventually" can be milliseconds (gossip in a healthy cluster) or hours (cross-region replication during a partition).
See: Simulators › cap theoremRead full entry
- FIDO2 / WebAuthn Security
Phishing-resistant authentication using public-key cryptography.
The browser API + the device protocol. Passkeys are FIDO2 credentials synced via cloud. The end of the password.
Read full entry- Forward secrecy Security
Past sessions stay secret even if the long-term key leaks.
Achieved with ephemeral key exchange (ECDHE). Mandatory in TLS 1.3 — RSA key transport is gone for this reason.
See: Guide › httpsRead full entry
- Garbage collection Concurrency
Automatic reclamation of unreachable memory.
Mark-sweep, generational, tri-color, concurrent. Java's G1, ZGC, Shenandoah; Go's tri-color concurrent collector; .NET's server GC.
See: Guide › garbage collectionRead full entry- GFS · Google File System Storage
The 2003 paper that founded scale-out storage.
Blueprint for HDFS, Colossus, every modern object store. Master + chunk servers; chunks are 64MB.
See: Papers › papersRead full entry- Gossip protocol Distributed
Each node periodically tells a random peer what it knows.
Information spreads in O(log N) rounds. Cassandra, Consul, Riak. No single coordinator to fail.
Read full entry- gRPC Networking
HTTP/2 + Protobuf RPC framework.
Bi-directional streaming, codegen in 12+ languages, schema-first. Default for internal microservice meshes.
See: Simulators › grpc vs restRead full entry
- Hash table Storage
O(1) average lookup via key → bucket index.
Hans Peter Luhn, IBM 1953. Collision strategies: chaining and open addressing. Every dictionary in every language.
See: DS › data structuresRead full entry- Hedged requests Performance
Send a duplicate request to a second server if the first is slow; take whichever returns first.
Dean & Barroso 2013. Cuts tail latency at modest cost. Tied requests cancel the loser.
See: Papers › papersRead full entry- High water mark · HW Distributed
Highest offset acknowledged by all in-sync replicas.
Kafka concept. Consumers can only read up to the HW; this is what makes the log durable. HW lags the leader by one round-trip.
See: Guide › kafkaRead full entry- Hinted handoff Distributed
When a replica is down, a temporary node holds its writes and replays them later.
Dynamo-style availability technique. Handoffs are time-bounded; if the replica stays down too long, anti-entropy takes over.
Read full entry- HLL · HyperLogLog Storage
Cardinality estimator in 12 KB with ~1% error.
Flajolet et al, INRIA 2007. Mergeable across shards. Powers every "unique users" dashboard at internet scale.
See: DS › data structuresRead full entry- HNSW · Hierarchical Navigable Small World Algorithms
Multi-layer graph index for approximate nearest-neighbour search.
Malkov & Yashunin, 2016. Default ANN index in pgvector, Qdrant, Weaviate, Milvus. State-of-the-art recall; high memory cost.
See: Simulators › vector embedding· Papers › papersRead full entry- HPACK Networking
Header compression for HTTP/2.
RFC 7541. Static + dynamic Huffman tables. A repeated 200-byte User-Agent header collapses to one byte after the first send.
See: Simulators › http2 streamsRead full entry- HSTS · HTTP Strict Transport Security Security
Tells the browser: "no plaintext, ever again, for this domain".
Stops downgrade attacks once cached. The HSTS preload list ships in browsers and closes the first-visit gap.
See: Guide › httpsRead full entry
- Idempotency Architecture
A request can be retried without changing the result.
GET, PUT, DELETE are idempotent in HTTP. POST is not — which is why payment APIs use idempotency keys to deduplicate retries.
Read full entry- Idempotency key Architecture
A client-supplied UUID that lets the server deduplicate retried POST requests.
Stripe popularised the pattern. Server stores (idempotency_key → response) for ~24h; second request with same key returns the same response.
Read full entry- In-Sync Replicas · ISR Distributed
Replicas caught up to the leader within a time bound.
Kafka. Falling behind for >30s evicts you; catching back up readmits. Producer durability scales with ISR size.
See: Guide › kafkaRead full entry- Information entropy ML
H(X) = −Σ p log p. Expected bits per sample.
Shannon, 1948. The compression ceiling for any source. Higher entropy = more uncertainty.
See: Foundations › foundations· Papers › papersRead full entry- IPv6 Networking
128-bit IP addresses; 2^96× more space than IPv4.
No more NAT (mostly). 30+ years of slow rollout; finally hit ~50% of US traffic in 2023.
Read full entry- Isolation level Storage
How much concurrent transactions can see of each other.
Read uncommitted < read committed < repeatable read < snapshot < serializable. Each blocks one anomaly the previous allowed.
See: Simulators › isolation levelsRead full entry
- Jitter Performance
Variance in latency or arrival rate.
A constant 100ms is fine; latency that swings 5–500ms is debilitating. Sources: GC pauses, network microbursts, scheduling delays.
Read full entry- JWT · JSON Web Token Security
Self-contained signed (or encrypted) token: header.payload.signature.
Ubiquitous in OAuth and OIDC. Verify offline; revoke through context (short TTL + refresh rotation).
See: Guide › oauth· Simulators › jwt lifecycleRead full entry
- Lamport timestamps Distributed
Logical counter per process; causal ordering of events.
Lamport 1978. Each event's timestamp is max(local, received) + 1. Two events with timestamps in different order may have any real-time ordering — but causally-ordered events keep their order.
See: Papers › papersRead full entry- Leader election Distributed
Picking exactly one node as the coordinator.
Paxos, Raft, ZooKeeper, etcd. Must tolerate failures and partitions without electing two leaders.
See: Simulators › raftRead full entry- Leader epoch Distributed
Monotonically-increasing term number for each elected leader.
Kafka KIP-101. New leader's first action is to fence out the old leader by epoch. Closes the "ghost record" data-loss window.
See: Guide › kafkaRead full entry- Little's Law Performance
L = λW. Items in system = arrival rate × average wait.
Holds for any stable queue regardless of distribution. The most-used theorem in operations research.
See: Foundations › foundationsRead full entry- LSM-tree · Log-Structured Merge-Tree Storage
Append-only writes flushed and merged in the background.
O'Neil et al, 1996. RocksDB, LevelDB, Cassandra, BigTable. Trades write amp for sequential I/O — SSDs love it.
See: Papers › papers· DS › data structuresRead full entry
- Memoization Algorithms
Cache function results by argument tuple.
Bottom of dynamic programming. Trivially turns exponential recursion into polynomial time, at the cost of memory.
Read full entry- Memory barrier Concurrency
Instruction that prevents reordering of memory operations across it.
CPUs and compilers reorder for performance. Concurrent code uses barriers to ensure visibility ordering. acquire / release / sequentially-consistent.
Read full entry- Memory hierarchy Performance
L1 ~1ns, L2 ~4ns, L3 ~12ns, RAM ~100ns, SSD ~100µs, network ~1ms.
Each level is ~10× slower than the last. Cache locality dominates algorithmic complexity for small n.
See: Foundations › foundationsRead full entry- Merkle tree Storage
Tree of hashes; root commits to all leaves.
Ralph Merkle, 1979. Bitcoin, Git, ZFS, Certificate Transparency, IPFS. Inclusion proofs are O(log n) hashes.
See: DS › data structuresRead full entry- MinHash Algorithms
Set similarity estimation via minimum hash matching.
Andrei Broder at AltaVista, 1997. Compute Jaccard similarity in O(k) instead of O(|A∪B|). Used in web deduplication.
See: DS › data structuresRead full entry- mTLS Security
Mutual TLS — both client and server present certificates.
Service mesh default for cross-service auth. Token-binding alternative is DPoP.
See: Guide › oauthRead full entry- MVCC · Multi-Version Concurrency Control Storage
Each row keeps versions; readers don't block writers.
Postgres, MySQL InnoDB, Oracle, CockroachDB. Vacuum/cleanup reclaims old versions.
Read full entry
- OAuth 2.1 Security
The modern auth-delegation flow: code + PKCE for everyone.
Removes implicit and password grants. Adds mandatory PKCE, exact-match redirect URIs, refresh token rotation.
See: Guide › oauthRead full entry- OCSP stapling Security
Server pre-fetches a signed revocation response and attaches it to the handshake.
Faster than client-side OCSP and doesn't leak browsing history to the CA.
See: Guide › httpsRead full entry- OIDC · OpenID Connect Security
Identity layer on top of OAuth 2.0 — adds an ID token (JWT).
Standardises "who is the user". Issued alongside the access token after auth code exchange.
See: Guide › oidcRead full entry- Optimistic concurrency Storage
Read freely; check for conflicts at commit; retry if conflict.
Postgres serializable level uses it (SSI). Better than locks under low contention; degrades to busy-waiting at high contention.
Read full entry- Outbox pattern Architecture
To atomically write to DB + publish a message: write to an outbox table in the same DB transaction.
A separate process polls the outbox and publishes. Solves "double-write" — without it, message+DB updates can't be atomic.
Read full entry
- p99 latency Performance
99th percentile response time.
For fan-out queries, system latency is dominated by the slowest. p99 of one becomes p50 of 100. Hedged requests cut the tail at modest cost.
See: Papers › papersRead full entry- Paxos Distributed
Family of protocols for distributed consensus.
Lamport, 1989. Notoriously dense; "Paxos Made Simple" is the path in. Variant Multi-Paxos powers Spanner, Chubby.
See: Papers › papersRead full entry- PKCE · Proof Key for Code Exchange Security
Cryptographic binding between authorisation request and token exchange.
RFC 7636. Stops authorization-code interception. Required for every OAuth 2.1 client (public and confidential).
See: Guide › oauthRead full entry- PoP · Point of Presence Networking
A CDN edge location.
Cloudflare has 300+. Closest to the user wins; anycast routing handles the picking.
See: Guide › cdnRead full entry
- QUIC Networking
UDP-based multiplexed and secure transport — the substrate of HTTP/3.
Per-stream loss recovery (no TCP HOL), 0-RTT resumption, integrated TLS 1.3, connection migration via connection ID.
See: Papers › papersRead full entry- Quorum Distributed
A majority needed for a write or read to count.
In a 5-node system, quorum is 3. Quorum reads + quorum writes = strong consistency. R+W>N is the math.
See: Simulators › read write quorumRead full entry
- Raft Distributed
Consensus algorithm explicitly designed to be teachable.
Ongaro & Ousterhout, 2014. Powers etcd, Consul, CockroachDB, TiKV. Easier to reason about than Paxos.
See: Simulators › raft· Papers › papersRead full entry- Rate limiter Performance
Caps request rate per key.
Token bucket, leaky bucket, fixed window, sliding window. Each has different burst tolerance.
See: Simulators › rate limiterRead full entry- Refresh token Security
Long-lived OAuth credential to mint new access tokens.
Modern: rotate on every use, revoke the chain on reuse detection. Storage matters — httpOnly cookie or OS keychain, never localStorage.
See: Guide › oauthRead full entry- Rendezvous hashing Algorithms
Alternative to consistent hashing — pick the node that scores highest for the key.
For each (key, node) pair compute hash; key goes to node with max hash. No virtual nodes needed; balanced by construction.
Read full entry- Reverse proxy Networking
Server-side proxy that fronts your backend.
NGINX, Envoy, Caddy, HAProxy. TLS termination, load balancing, caching, header rewriting.
See: Guide › reverse proxyRead full entry- RPO / RTO Operations
Recovery Point / Recovery Time Objective.
RPO = how much data you can lose (e.g. 5 min). RTO = how long the recovery takes (e.g. 1 hour). Drives backup frequency and replication topology.
Read full entry
- Saga pattern Architecture
Long-running distributed transaction as a sequence of compensable steps.
Each step has a corresponding compensation. Replaces 2PC for cross-service transactions where atomicity isn't feasible.
Read full entry- Sharding Storage
Splitting data across machines by key.
Range, hash, directory. The harder problems are rebalancing, cross-shard queries, hot keys.
See: Simulators › database shardingRead full entry- Skip list Storage
Probabilistic balanced search structure.
William Pugh, 1989. Tall nodes form an "express lane". Lock-free skip lists are easier to write than lock-free balanced trees — why Redis ZSET uses them.
See: DS › data structuresRead full entry- SLO · Service Level Objective Operations
A target for an SLI (latency, availability).
Internal commitment, looser than SLA. Tied to error budget — burn it slowly or roll back.
Read full entry- Snapshot isolation Storage
Each transaction sees a consistent snapshot taken at start.
Postgres, Oracle default. Prevents dirty reads, non-repeatable reads, phantoms — but allows write-skew anomalies.
See: Simulators › isolation levelsRead full entry- SNI · Server Name Indication Security
TLS extension that tells the server which hostname is being requested.
Lets one IP host many TLS sites. Travels in the clear; ECH (Encrypted Client Hello) hides it.
See: Guide › sni· Guide › httpsRead full entry- Speed of light Performance
~3×10⁸ m/s vacuum, ~2×10⁸ m/s in fibre.
New York → London round-trip: ~56ms minimum. The floor under every cross-region replication.
See: Foundations › foundationsRead full entry- Spinlock Concurrency
A lock implemented with a busy-wait loop instead of going to sleep.
Right when contention is rare and critical sections are short (microseconds). Wrong almost everywhere else — burns CPU.
Read full entry- Stream Networking
A logical flow of frames over a single connection.
HTTP/2 introduced streams to multiplex requests. HTTP/3 made each stream independent at the transport layer (no TCP HOL).
See: Simulators › http2 streamsRead full entry
- Tail latency Performance
The slow end of the distribution — p95, p99, p99.9.
Dean & Barroso, 2013. For fan-out systems, the tail dominates. Hedged requests, redundant requests, soft timeouts mitigate.
See: Papers › papersRead full entry- TCP slow-start Networking
Begin with a small congestion window; double per RTT until loss.
Van Jacobson, 1988. The reason new connections are slow until they're warmed up; why HTTP keep-alive matters.
See: Simulators › tcp congestionRead full entry- Time-to-live · TTL Networking
A counter that limits how long something stays valid.
IP packets (hop count, used by traceroute), DNS records (cache duration), cache entries (eviction).
See: Guide › dnsRead full entry- TLS 1.3 Security
2018 redesign of Transport Layer Security.
1-RTT handshake, AEAD-only ciphers, mandatory ECDHE, PSK resumption. The standard.
See: Guide › https· Simulators › tls handshakeRead full entry- Token bucket Performance
Rate limiter that accumulates tokens at a fixed rate.
Allows bursts up to bucket size, smooth rate at long term. Default for most rate limits in cloud APIs.
See: Simulators › rate limiterRead full entry- Transaction log Storage
Append-only record of every change for crash recovery.
Same idea as WAL. Postgres pg_wal, MySQL redo log, SQL Server transaction log. Replay on crash to recover committed transactions.
See: Guide › walRead full entry- TrueTime Distributed
Google's globally-synchronised clock with a known uncertainty bound.
Hardware GPS + atomic clocks. Lets Spanner do externally-consistent transactions globally — wait out the uncertainty before committing.
See: Papers › papersRead full entry- Two-phase commit · 2PC Distributed
Coordinator-driven atomic commit across N participants.
Prepare → commit. Blocking: if the coordinator dies after prepare, participants are stuck. Saga and Calvin are alternatives.
Read full entry
- Union-Find Algorithms
Track connected components in O(α(n)) ≈ O(1).
Galler & Fischer, 1964; Tarjan's analysis 1975. Path compression + union-by-rank. Powers Kruskal MST and equivalence-class problems.
See: DS › data structuresRead full entry- Universal Scalability Law · USL Performance
C(N) = N / (1 + α(N−1) + βN(N−1)). Throughput vs concurrency.
Above a critical N, adding workers slows throughput. β captures coordination cost (O(N²)). The crisper version of "Amdahl plus contention plus coherence".
See: Foundations › foundationsRead full entry
- WAL · Write-Ahead Log Storage
Append every change to a log before applying it to data files.
Crash recovery: replay the log. The basis for every modern database's durability story.
See: Guide › walRead full entry- WebSocket Networking
Persistent bi-directional connection over a single TCP socket.
Upgrades from HTTP via 101 Switching Protocols. The default for browser-side real-time before HTTP/3 streams.
See: Guide › websocketsRead full entry- Wide-column store Storage
Schema is dynamic per row; many sparse columns.
Bigtable, HBase, Cassandra, ScyllaDB. Optimised for time-series and event data where most rows have most columns null.
See: Papers › papersRead full entry- Work stealing Concurrency
Idle workers steal pending tasks from busy workers' queues.
Each worker has a deque; thieves take from the back, owner from the front. Java ForkJoinPool, Go scheduler, Rust Rayon, .NET TPL.
See: Simulators › goroutine schedulerRead full entry
From a term, to its source.
Every entry links to where the term is explained at length. Start with what you don't know; follow the link to what explains it.