The failure: a healthy socket can carry an incomplete story

An exchange connection can reopen after an interruption while the local order book still lacks the updates missed during the gap. A connection-health indicator answers whether transport is open. It does not answer whether a reconstructed book is valid.

Market-data failure investigation therefore needs more than reconnect counters. It must examine source progress, reconstruction state, event handling and durable persistence. This note combines a documented source-continuity contract with a separate reproduced queue failure; it does not claim both explain every historical interruption.

Investigation: distinguish silence from local delay

A venue may stop sending relevant updates. A local event loop may also delay callbacks even while the underlying connection remains open. Both can look like a source going quiet unless the diagnostics distinguish the stages.

In a documented burst investigation, repeated removal from the front of an array reproduced a long synchronous drain. That blocked opportunities for socket and timer callbacks. The report identified it as a root contributor, without attributing every millisecond of the historical event stall to that one function.

Implemented changes

The queue remediation uses ring storage and bounded cooperative draining. Consumed references are released, FIFO is retained, and yielding gives the event loop another opportunity to service I/O. This improves the reproduced queue behavior without changing factual timestamps or claiming a hard real-time guarantee.

Separately, the documented order-book continuity path invalidates trust on disconnect, retires the old socket and starts a new recovery cycle. Depending on the venue, recovery obtains a new REST or socket snapshot before accepting a valid continuation. Messages from retired connections cannot restore trust.

Verification: connection, continuity and landing

The queue report compares preserved pre-fix and revised implementations with a deterministic backlog. Its checks cover input/output counts, FIFO, duplicates and callback opportunities. Persistence is represented by an I/O yield in that focused benchmark, so the result is not end-to-end Parquet throughput.

The source-continuity contract also checks actual book progress, time spent trusted or untrusted, fresh data near orderly shutdown and the requested acquisition duration. Repeated socket opens alone do not satisfy the freshness condition.

  • Transport: is the connection open?
  • Continuity: is the reconstructed state valid?
  • Progress: are relevant source frames arriving?
  • Durability: has accepted data reached confirmed storage?
  • Coverage: did every required lane satisfy the run contract?

Remaining limitations

A fresh snapshot can establish a new valid starting point, but it does not recreate the missing interval. Replay must retain the interruption instead of presenting a continuous history.

Focused regression checks support specific code paths. They cannot prove the absence of every future source outage, scheduling delay or persistence failure. Data completeness remains a property to verify for each run and lane.

Evidence basis

Based on the private project’s operational Level-2 source-continuity documentation and September 2026 temporal-burst investigation. No raw market data, private source code, network endpoints or operational configuration is reproduced.