Consistency Models Anomaly Lab (Interactive)
Run scripted read probes against five models and watch anomalies appear and disappear. Strong, sequential, causal, read-your-writes, and eventual consistency probed with post-then-reply scenarios that expose exactly which guarantee each model drops.
Consistency Spectrum Anomaly Probe
A post and its reply race down two replication paths. Replay three observer reads at t=60/80/120ms under each model and count the anomalies.
Eventual: Zero coordination; anomalies until background sync. Push the lag slider to 0 to watch every model momentarily agree — consistency strength is about what happens when they do not.
How It Works Under the Hood
Consistency is not binary: it is a spectrum of guarantees each defined by the anomalies it forbids. Strong consistency forbids reading old values at all; sequential keeps every observer seeing one global order; causal preserves dependencies between writes; read-your-writes only promises you never lose sight of your own edits; eventual promises convergence with no window bound. This lab replays a fixed write story — a post, then a reply that cites it — against your chosen replica model, replication lag, and partition state, then reports which guarantees the probe scenario violated.
Core Architectural Principles
- Deterministic probe scenario tests read-your-writes, monotonic reads, and causal visibility.
- Replication lag slider decides when each anomaly actually fires under weak models.
- CP-tier models switch from stale answers to hard errors when the partition toggle is set.
Never say “eventually consistent” alone — name the property you need: causal for threads, read-your-writes for user edits, monotonic reads so counters never go backwards. Then give a convergence bound: “staleness bounded by replication lag, p99 under 1 second,” which is the answer reviewers want.
Each stronger guarantee you add removes a real anomaly but hands back a slice of latency or availability.