Truthound · Streams and change data · 1/2
Examine retries, ordering and checkpoints in result sets.
29 articles
Winlog rollover recovery: verify live intake and EVTX backlog separately
New live events do not establish recovery of unread archive records.
Debezium PostgreSQL TOAST arrays: do not overwrite an unavailable field with an empty list
Separate an unavailable TOAST array from an actual change to an empty array, then validate the sink’s overwrite decision.
SQL Server when_needed: do not accept a NULL LSN as a recovery position
Separate an offset record from a usable LSN and verify the required snapshot actually starts.
Spanner change streams: an RPC end is different from restriction completion
Distinguish an artificial end of one RPC from completion of the whole restriction.
PostgreSQL publication columns shifted: reconcile the interval before restart
A restart can refresh Relation metadata without correcting rows decoded earlier.
Logstash HTTP pipelining: method locks do not establish record assembly order
Protecting individual methods is different from preserving complete-request order.
A missing Logstash DLQ segment: check the next writer before deleting data
The impact of the same exception depends on its position in the writer lifecycle.
Debezium IBM i snapshots: did the initial load cover every member that CDC can read?
Reconcile source and snapshot data by member. Matching row counts still require matching keys, duplicate checks and a common reference time.
IBM i blocking snapshots: verify the pause handshake after request acceptance
Request acceptance, streaming pause acknowledgment and snapshot execution are separate states.
B@20 after a Flink checkpoint: a watermark can pass the remaining buffer
A completed checkpoint buffer is different from the current buffer receiving later records.
Filebeat cursor cache: can receiver startup order change the state index?
Input identity and cached-store ownership must be checked separately.
ETW EventsLost=0 but fewer records arrive: reconcile independent event keys
Zero loss counters do not establish complete capture.
PostgreSQL decoderbufs resume: a BEGIN can consume a row skip
Count control messages separately from rows sharing an LSN.
Bigtable RESUME_OR_NEW: count partition consumers inside one job
Job count is different from partition-consumer ownership.
Beam YAML 1d: align the day label with an actual 86,400-second window
A fixed UTC day requires 86,400 seconds; a configuration label alone does not establish that duration.
Beam OrderedListState: retain both A and B at the same timestamp
A sorting timestamp does not replace event identity; check bucket counts and the actual members.
Beam Parquet window mixing: readable files still need membership checks
Compare event identity with its assigned window across interleaving and flush boundaries.
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.
Keep nested delete ordering values in Hudi event-time reconciliation
Replay a late delete against a newer update for the same key. Record the nested ordering path and table version, then check the survivor against the event-time policy.
Keep net-zero rewrites from erasing a Doris offset snapshot key
A same-value rewrite can still require the earlier row in an offset-based snapshot. Test the effect of net-delta elimination against keys fixed at that offset while preserving the report's unexecuted-scenario scope.
