M3 · Best Effort Consistency
The second strategy on the menu, and the deliberate opposite of the first: guarantee a response always, and provide consistency as best the network allows — the right choice for content served to readers worldwide, where a slightly stale answer is useful and a withheld one is not.
The classic example is a content delivery network caching close to users; echd takes the same stance on its market-data feed and its globally replicated retention — answering from local state and bounding staleness rather than forbidding it, marking a value stale under partition instead of refusing it. Where the consistency-first ledger of M2 blocks rather than diverges, these surfaces diverge within a bound rather than block. The strategy and its examples are settled and external (Gilbert & Lynch, Akamai, Valkey, Tigris); echd's composition is what this course proposes.
state · hover to walk the near-cache In the common case the reader serves the most recent value from local memory with no round-trip; when the invalidation channel is cut by a partition, it serves the value marked stale rather than withholding it. Hover a state below to take it one move at a time.
Hover or focus a state to walk it; click to pin. Without scripting, all four read in sequence and the figure shows every marker at once.
Best-effort consistency makes the opposite commitment to best-effort availability: a response is guaranteed, and consistency is provided as well as it can be without ever blocking. It fits where two conditions hold — a stale answer is tolerable for a while, and a stopped answer is not — which is the case for content served to a wide, geographically spread audience.
Its classic example is a content delivery network: a CDN caches content at edge servers close to end users to avoid middle-mile bottlenecks, serving from the nearest cache and refreshing on a time-to-live, so a reader may see a value bounded in staleness by the TTL. Naming served through caching, where a resolver answers from a cached record until it expires, is the same pattern. The lesson mirrors the first chapter: availability-first is not a refusal to be correct — it is a decision to answer now and reconcile toward correctness, choosing availability over consistency when the two collide.
The commitment
A response is guaranteed; consistency is provided as well as the network allows, and never traded for a block. The opposite of the consistency-first surface, which blocks rather than diverges.
silence is the failure · availability over consistency
The bound
A reader may see a value bounded in staleness by the TTL or by the partition's duration — divergence within a bound rather than a withheld answer. The dial that tunes the bound is M4.
bounded staleness · diverge within a bound
The example
Content cached at edge servers close to readers, served from the nearest cache, refreshed on a TTL. Naming served through caching is the same pattern — both Gilbert and Lynch's catalogue entry.
cache close, refresh on TTL · the classic example
Retention and delivery are the other availability-first surfaces, and echd splits them across two tiers. Each instrument has its own append-only stream — a log with auto-generated, time-ordered IDs, appended with one command and tailed by readers or shared across a consumer group, with old entries trimmed to keep memory bounded. The stream is the ordered, replicated record that readers consume without contending with the writer.
Deep retention runs on Tigris, a globally distributed, S3-compatible object store that keeps an object close to the region where it was written and replicates it close to the readers who request it, giving low-latency global reads and replacing a CDN with no configuration to manage. Both tiers answer from local or nearby state and accept bounded staleness as the price of always answering.
Hover or focus a tier to read it; click to pin. Without scripting, all three read in sequence.
The third dive, M3.3, builds the per-instrument streams and the global edge — the retention and delivery surfaces that answer from nearby state and bound their staleness rather than block.
This landing states the chapter's argument once; each dive takes one movement to the depth a comprehensive treatment needs.
The strategy and its classic example: a content cache spread worldwide, always responsive, consistent in a best-effort sense.
M3.2The near-cache and coherence broadcast; serving the most recent value; marked stale under partition.
M3.3Per-instrument streams and globally replicated Tigris storage as availability-first retention and delivery.
CAP made rigorous: the safety-versus-liveness framing, the two-group proof, and the menu that replaces “two of three” with a per-surface design question.
the proof both strategies answer to — this chapter's surfaces choose liveness where M2's choose safety.
/bcsThe law in three clauses and the 14-byte branded-snowflake identity contract, proven by frozen transcripts across five runtimes.
the identity that keys a cached value and a stream entry rests on the contract built here.
Sources
Related