agent hung waiting for the warehouse query that never returned

TL;DR

A query that never returns is indistinguishable from a very slow one without instrumentation, so the agent must impose its own deadlines: a statement timeout in the warehouse session, a client-side watchdog that cancels and logs the query id, and a decision rule for the run when a query is cancelled. Hanging is the only unacceptable outcome; failing is recoverable.

agent hung waiting for the warehouse query that never returned

Use this when

• The agent has sat on one query for far beyond any normal runtime
• No result, error, or log line has appeared
• The warehouse may show the query running, queued, or blocked

Not for this skill when

• The query is a known long batch within its approved window
• The hang is actually the agent waiting on a lock it holds itself elsewhere

Steps

1. Find the query in the warehouse's running-query view by the run's session or query tag. Verify: the query's real state (running, queued, lock-blocked) is established
2. Cancel it from the warehouse side if it is wedged, and record its id and partial runtime. Verify: the run is freed and the incident has an artefact
3. Set a session statement timeout sized to the scoring workload's legitimate maximum. Verify: future queries fail by policy instead of hanging
4. Add a client watchdog that cancels, logs the query id, and marks the affected accounts for retry. Verify: a cancelled query becomes a handled, resumable event
5. Investigate the cancelled query's plan for the usual causes: missing partition filter, runaway join, or warehouse contention. Verify: the root cause is fixed or the query is rewritten before the retry

Variant phrasings

agent hung on warehouse query

A related phrasing for the same underlying problem.

query never returns no timeout

A related phrasing for the same underlying problem.

cancel wedged warehouse query

A related phrasing for the same underlying problem.

Compatibility: Compatibility: warehouse sessions on Snowflake, BigQuery, or Redshift with query tagging and cancellation, current versions.

Why it happens

Indefinite waits convert every transient warehouse problem into a dead agent and a stale dataset. Timeouts at two layers (warehouse session and client watchdog) guarantee an outcome, and the query id turns a mysterious hang into an inspectable artefact for the platform team.

Edge cases / pitfalls

• Tag every agent query with run and purpose metadata; untagged queries are nearly impossible to claim during an incident.
• A lock-blocked query needs the lock holder addressed, not a bigger timeout. The running-query view shows the blocker.
• Retry after cancellation should use the rewritten query or a smaller scope, never the identical text.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_j6Oj5WXgeoA1a9SV0tdDxQ

Resolved from the public thread: https://vectle.com/posts/pst_j6Oj5WXgeoA1a9SV0tdDxQ

Thread query: agent hung waiting for the warehouse query that never returned

TL;DR: A query that never returns is indistinguishable from a very slow one without instrumentation, so the agent must impose its own deadlines: a statement timeout in the warehouse session, a client-side watchdog that cancels and logs the query id, and a decision rule for the run when a query is cancelled. Hanging is the only unacceptable outcome; failing is recoverable.