Module evidence 

What a soak run records, in the form the verdict is computed from.

The harness writes one evidence.json at the end of a run, and appends each fault to faults.jsonl as it is injected and cleared, so a run that dies mid-way still leaves its fault timeline behind. Every timestamp is wall-clock Unix milliseconds, the coordinate system the broker’s own logs and metrics use, so the fault record lines up with them.

Records are identified by (producer, seq): each producer numbers what it sends from zero and writes the number into the record value, so a reader can say which send a record came from without trusting offsets that a failover may have reassigned.

Structs

CommitCheck

A committed offset checked across a fault.

DuplicateRead

One record read at more than one position.

Evidence

Everything one run recorded.

FaultRecord

A fault as it was executed.

InjectedLoss

One record a self-check deleted from the cluster.

LatencySummary

Latency percentiles over one window kind.

PathMeasurement

One path’s measured latency and throughput.

ProducerLedger

One producer’s account of what it sent.

ReaderLedger

What one reader saw of one producer’s records, reading from the earliest offset after the load stopped.

RecoveryRecord

One recovery, measured after a fault.

TxnLedger

One transaction and how it ended.

Enums

TxnOutcome

How a transaction ended, as its producer saw it.

Constants

FINAL_COMMIT_CHECK

The fault of the committed-offset check taken after the load stops.

Functions

parse_record_tag

The (producer, seq) a record value carries, or None for a value that no soak producer wrote.

record_tag

The record value prefix that carries a record’s identity: <producer>/<seq>;. The rest of the value is padding.