[
https://issues.apache.org/jira/browse/SPARK-56537?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18077581#comment-18077581
]
Dhruv Patel edited comment on SPARK-56537 at 5/2/26 2:42 AM:
-------------------------------------------------------------
Thanks [~kabhwan]! Did the audit -- summary below.
h3. Bug surface in \{{StateOperatorProgress}}
Three per-batch time fields are never reset on the no-data path:
allUpdatesTimeMs, allRemovalsTimeMs, commitTimeMs.
The other three per-batch fields (numRowsUpdated, numRowsRemoved,
numRowsDroppedByWatermark) are already reset. The five snapshot fields
(numRowsTotal, memoryUsedBytes, numShufflePartitions, numStateStoreInstances,
operatorName) should keep their values, as today.
h3. \{{customMetrics}} splits across two registries
The trait subtypes already encode most of it cleanly.
h4. (A) \{{StateStoreCustomMetric}} (provider-level)
* \{{StateStoreCustomTimingMetric}} -- always per-batch
* \{{StateStoreCustomSumMetric}} -- always per-batch
* \{{StateStoreCustomSizeMetric}} -- mostly per-batch byte/file counters, with
a small snapshot subset
Snapshot Size metrics (the only ambiguous group, 6 names total):
* *RocksDB (5):* rocksdbSstFileSize, rocksdbPinnedBlocksMemoryUsage,
rocksdbNumInternalColFamiliesKeys, rocksdbNumExternalColumnFamilies,
rocksdbNumInternalColumnFamilies
* *HDFSBackedStateStoreProvider (1):* stateOnCurrentVersionSizeBytes
Provider tally for completeness:
* \{{RocksDBStateStoreProvider}}: 36 custom metrics (31 per-batch / 5 snapshot
Size) + 1 instance metric (StateStoreSnapshotLastUploadInstanceMetric,
ignoreIfUnchanged=false, max-combine -- snapshot-style).
* \{{HDFSBackedStateStoreProvider}}: 5 custom metrics (4 Sum per-batch / 1
snapshot Size) + 1 instance metric (same
StateStoreSnapshotLastUploadInstanceMetric).
h4. (B) \{{StatefulOperatorCustomMetric}} (operator-level)
14-15 metrics across 3 operators, all StatefulOperatorCustomSumMetric, all
per-batch by mechanism (verified by tracing every increment site):
* \{{BaseStreamingDeduplicateExec}}: numDroppedDuplicateRows.
* \{{StreamingSymmetricHashJoinExec}}: skippedNullValueCount (only declared
when storeConf.skipNullsForStreamStreamJoins=true).
* \{{TransformWithStateExecBase}}: 13 metrics -- initialStateProcessingTimeMs,
numValueStateVars, numListStateVars, numMapStateVars, numDeletedStateVars,
timerProcessingTimeMs, numRegisteredTimers, numDeletedTimers, numExpiredTimers,
numValueStateWithTTLVars, numListStateWithTTLVars, numMapStateWithTTLVars,
numValuesRemovedDueToTTLExpiry. All incremented inside batch processing
(StatefulProcessorHandleImpl, TransformWithStateExec) -- none are snapshot
reads.
{panel:title=Worth flagging on
TransformWithState|borderStyle=dashed|borderColor=#ccc|bgColor=#f4f5f7}
The six numXxxStateVars and numXxxStateWithTTLVars metrics are mechanically
per-batch (incremented per getValueState / getListState / getMapState call) but
a user may interpret them as snapshots ("how many state vars does my query
have"). Resetting them to 0 on idle triggers means the dashboard value flickers
between batches. If you'd rather preserve them, that suggests they want
snapshot semantics and the operator should be setting them differently --
separate discussion. For SPARK-56537 I'd default to "reset (since they are
mechanically per-batch)" unless you say otherwise.
{panel}
No other stateful operators (StateStoreSaveExec, StateStoreRestoreExec,
FlatMapGroupsWithStateExec, FlatMapGroupsInPandasWithStateExec,
StreamingGlobalLimitExec, SessionWindowStateStoreRestoreExec,
EventTimeWatermarkExec) override customStatefulOperatorMetrics.
SessionWindowStateStoreSaveExec.getProgress overrides only to thread
numRowsDroppedByWatermark from the upstream restore exec; the returned
StateOperatorProgress still flows through resetExecStatsForNoExecution normally.
h3. Edge cases confirmed handled today
lastExecution == null first-trigger case (guarded at ProgressReporter:602),
multi-operator queries (resetExecStatsForNoExecution already iterates per
operator), SessionWindowStateStoreSaveExec override (no bypass), Spark UI
(reads fields verbatim, benefits for free), protobuf serializer (pure
passthrough, no changes needed).
h3. Proposed plan -- want to confirm
# Reset allUpdatesTimeMs, allRemovalsTimeMs, commitTimeMs unconditionally on
the no-data path. Straightforward.
# For customMetrics, reset every entry whose underlying metric is
StateStoreCustomTimingMetric, StateStoreCustomSumMetric, or
StatefulOperatorCustomSumMetric (all unambiguously per-batch). For
StateStoreCustomSizeMetric, reset every entry except the six-name snapshot
whitelist above. Two implementations:
#* *(a)* Hardcode the six snapshot Size names in resetExecStatsForNoExecution.
Small whitelist, low maintenance burden, easy to read.
#* *(b)* Extend StateStoreCustomMetric with isSnapshot: Boolean (default
false). Each provider declares semantics at the source. Slightly bigger
surface, future-proof.
#* Given the snapshot subset is only six names, (a) feels right. Preference?
# Confirm TransformWithState's numXxxStateVars / numXxxStateWithTTLVars
semantics -- reset (treat as per-batch, my default) or preserve (treat as
user-visible snapshot, would be inconsistent with their type)?
# StateStoreInstanceMetric out of scope -- both providers'
StateStoreSnapshotLastUploadInstanceMetric is monotonic-max-combine and
snapshot-style, so leaving as-is is correct.
# *API shape:* leaning toward adding resetForNoExecution() on
StateOperatorProgress rather than growing copy()'s arg list further (3 args
after SPARK-56464, would become 6+ with this fix). Co-locates the reset logic
with the data class. Open to either.
Master only as you noted, no backport. If you don't have strong opinions on
(2), (3), or (5), I'll go with (a) + reset-the-TWS-state-var-metrics +
resetForNoExecution(), add tests in the new ProgressReporterSuite (extending
the existing test plus one for customMetrics on RocksDB), and ping you on the
PR.
was (Author: JIRAUSER312625):
Thanks [~kabhwan]! Did the audit -- summary below.
h3. Bug surface in \{{StateOperatorProgress}}
Three per-batch time fields are never reset on the no-data path:
\{{allUpdatesTimeMs}}, \{{allRemovalsTimeMs}}, \{{commitTimeMs}}.
The other three per-batch fields (\{{numRowsUpdated}}, \{{numRowsRemoved}},
\{{numRowsDroppedByWatermark}}) are already reset. The five snapshot fields
(\{{numRowsTotal}}, \{{memoryUsedBytes}}, \{{numShufflePartitions}},
\{{numStateStoreInstances}}, \{{operatorName}}) should keep their values, as
today.
h3. \{{customMetrics}} splits across two registries
The trait subtypes already encode most of it cleanly.
h4. (A) \{{StateStoreCustomMetric}} (provider-level)
* \{{StateStoreCustomTimingMetric}} -- always per-batch
* \{{StateStoreCustomSumMetric}} -- always per-batch
* \{{StateStoreCustomSizeMetric}} -- mostly per-batch byte/file counters, with
a small snapshot subset
Snapshot Size metrics (the only ambiguous group, 6 names total):
* *RocksDB (5):* \{{rocksdbSstFileSize}}, \{{rocksdbPinnedBlocksMemoryUsage}},
\{{rocksdbNumInternalColFamiliesKeys}}, \{{rocksdbNumExternalColumnFamilies}},
\{{rocksdbNumInternalColumnFamilies}}
* *HDFSBackedStateStoreProvider (1):* \{{stateOnCurrentVersionSizeBytes}}
Provider tally for completeness:
* \{{RocksDBStateStoreProvider}}: 36 custom metrics (31 per-batch / 5 snapshot
Size) + 1 instance metric (\{{StateStoreSnapshotLastUploadInstanceMetric}},
\{{ignoreIfUnchanged=false}}, max-combine -- snapshot-style).
* \{{HDFSBackedStateStoreProvider}}: 5 custom metrics (4 Sum per-batch / 1
snapshot Size) + 1 instance metric (same
\{{StateStoreSnapshotLastUploadInstanceMetric}}).
h4. (B) \{{StatefulOperatorCustomMetric}} (operator-level)
14-15 metrics across 3 operators, all \{{StatefulOperatorCustomSumMetric}}, all
per-batch by mechanism (verified by tracing every increment site):
* \{{BaseStreamingDeduplicateExec}}: \{{numDroppedDuplicateRows}}.
* \{{StreamingSymmetricHashJoinExec}}: \{{skippedNullValueCount}} (only
declared when \{{storeConf.skipNullsForStreamStreamJoins=true}}).
* \{{TransformWithStateExecBase}}: 13 metrics --
\{{initialStateProcessingTimeMs}}, \{{numValueStateVars}},
\{{numListStateVars}}, \{{numMapStateVars}}, \{{numDeletedStateVars}},
\{{timerProcessingTimeMs}}, \{{numRegisteredTimers}}, \{{numDeletedTimers}},
\{{numExpiredTimers}}, \{{numValueStateWithTTLVars}},
\{{numListStateWithTTLVars}}, \{{numMapStateWithTTLVars}},
\{{numValuesRemovedDueToTTLExpiry}}. All incremented inside batch processing
(\{{StatefulProcessorHandleImpl}}, \{{TransformWithStateExec}}) -- none are
snapshot reads.
{panel:title=Worth flagging on
TransformWithState|borderStyle=dashed|borderColor=#ccc|bgColor=#f4f5f7}
The six \{{numXxxStateVars}} and \{{numXxxStateWithTTLVars}} metrics are
mechanically per-batch (incremented per
\{{getValueState}}/\{{getListState}}/\{{getMapState}} call) but a user may
interpret them as snapshots ("how many state vars does my query have").
Resetting them to 0 on idle triggers means the dashboard value flickers between
batches. If you'd rather preserve them, that suggests they want snapshot
semantics and the operator should be setting them differently -- separate
discussion. For SPARK-56537 I'd default to "reset (since they are mechanically
per-batch)" unless you say otherwise.
{panel}
No other stateful operators (\{{StateStoreSaveExec}},
\{{StateStoreRestoreExec}}, \{{FlatMapGroupsWithStateExec}},
\{{FlatMapGroupsInPandasWithStateExec}}, \{{StreamingGlobalLimitExec}},
\{{SessionWindowStateStoreRestoreExec}}, \{{EventTimeWatermarkExec}}) override
\{{customStatefulOperatorMetrics}}.
\{{SessionWindowStateStoreSaveExec.getProgress}} overrides only to thread
\{{numRowsDroppedByWatermark}} from the upstream restore exec; the returned
\{{StateOperatorProgress}} still flows through
\{{resetExecStatsForNoExecution}} normally.
h3. Edge cases confirmed handled today
{\{lastExecution == null}} first-trigger case (guarded at
\{{ProgressReporter:602}}), multi-operator queries
(\{{resetExecStatsForNoExecution}} already iterates per operator),
\{{SessionWindowStateStoreSaveExec}} override (no bypass), Spark UI (reads
fields verbatim, benefits for free), protobuf serializer (pure passthrough, no
changes needed).
h3. Proposed plan -- want to confirm
# Reset \{{allUpdatesTimeMs}}, \{{allRemovalsTimeMs}}, \{{commitTimeMs}}
unconditionally on the no-data path. Straightforward.
# For \{{customMetrics}}, reset every entry whose underlying metric is
\{{StateStoreCustomTimingMetric}}, \{{StateStoreCustomSumMetric}}, or
\{{StatefulOperatorCustomSumMetric}} (all unambiguously per-batch). For
\{{StateStoreCustomSizeMetric}}, reset every entry except the six-name snapshot
whitelist above. Two implementations:
#* *(a)* Hardcode the six snapshot Size names in
\{{resetExecStatsForNoExecution}}. Small whitelist, low maintenance burden,
easy to read.
#* *(b)* Extend \{{StateStoreCustomMetric}} with \{{isSnapshot: Boolean}}
(default \{{false}}). Each provider declares semantics at the source. Slightly
bigger surface, future-proof.
#* Given the snapshot subset is only six names, (a) feels right. Preference?
# Confirm TransformWithState's \{{numXxxStateVars}}/\{{numXxxStateWithTTLVars}}
semantics -- reset (treat as per-batch, my default) or preserve (treat as
user-visible snapshot, would be inconsistent with their type)?
# \{{StateStoreInstanceMetric}} out of scope -- both providers'
\{{StateStoreSnapshotLastUploadInstanceMetric}} is monotonic-max-combine and
snapshot-style, so leaving as-is is correct.
# *API shape:* leaning toward adding \{{resetForNoExecution()}} on
\{{StateOperatorProgress}} rather than growing \{{copy()}}'s arg list further
(3 args after SPARK-56464, would become 6+ with this fix). Co-locates the reset
logic with the data class. Open to either.
Master only as you noted, no backport. If you don't have strong opinions on
(2), (3), or (5), I'll go with (a) + reset-the-TWS-state-var-metrics +
\{{resetForNoExecution()}}, add tests in the new \{{ProgressReporterSuite}}
(extending the existing test plus one for \{{customMetrics}} on RocksDB), and
ping you on the PR.
> Some stateful operator metrics are not reset in no batch trigger progress
> event
> -------------------------------------------------------------------------------
>
> Key: SPARK-56537
> URL: https://issues.apache.org/jira/browse/SPARK-56537
> Project: Spark
> Issue Type: Bug
> Components: Structured Streaming
> Affects Versions: 4.2.0
> Reporter: Jungtaek Lim
> Priority: Minor
>
> [https://github.com/apache/spark/pull/55331#discussion_r3103933756]
> We reset several metrics (about num rows) on stateful operator metrics for no
> batch trigger progress event, but still we have more metrics to be audited.
> This ticket tracks the effort.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]