Styxis
Company
Knowledge

Require repeated progress from combined Iceberg maintenance tasks

Check repeated progress after the first completion when maintenance tasks share a lock. Count completions by task in a window based on the actual schedule to distinguish a stall from work that is not yet due.

Acceptance includes the next cycle

A recurring service is not accepted by its first successful run alone. Each task sharing the lock must reach the next required cycle. Truthound Core can check per-task completion counts in an operator-defined observation window. [1, 2, 3]

A combined-maintenance report

Iceberg issue #18435 reports ConvertEqualityDeletes combined with other Flink maintenance tasks on 1.12.0 failing to forward another task’s watermark, leaving the shared lock unreleased. The converter-only control succeeds. Checkpoint-restoration effects remain an untested hypothesis in the source. [1]

Join trigger and release epochs

Link task type, trigger epoch, completion epoch, release watermark, and count. Observe the combined configuration through each task’s second approved cycle. Derive the observation window from the real schedules so a task is not declared stuck before it should run. [1, 3]

The second-completion boundary

Declare the minimum illustrative boundary as 2 completed cycles per task. A count of 2 passes; 1 fails despite an initial success. This threshold tests continued progress and is not a universal maintenance service-level objective. [1, 3]

Counts do not prove lock cause

BetweenValidator in Truthound Core Python SDK 3.3.2 can check a prepared completion-count column with min_value=2. The operator supplies task identity and the window. The SDK does not forward Flink watermarks, release locks, or prove the cause from a count alone. [1, 2, 3]

Compare isolated and combined tasks

Compare converter-only and combined configurations under the same load and retain watermark and lock traces for the task that stops advancing. A successful isolated component is only one control in that investigation. [1, 3]

Approve the observed task set

Publish recurring readiness only for the task set actually observed. Completion counts establish the bounded progress condition; they do not certify all future scheduling or checkpoint recovery behavior. [1, 3]

Application conditions

Prepare an actual-schedule window and per-task trigger, completion, and release records.

Alternatives

Retain isolated/combined controls under the same load and watermark traces.

Limitations

  • Completion count establishes neither lock cause nor future checkpoint recovery.

Evidence

Event checked
Subject checked
Modified