prodriguezdefino opened a new pull request, #41072:
URL: https://github.com/apache/beam/pull/41072
Problem:
In Stage.Execute (sdks/go/pkg/beam/runners/prism/internal/stage.go), Prism
records final bundle metrics before requesting metadata for unknown short IDs
from the SDK worker:
unknownIDs := j.ContributeFinalMetrics(resp)
if len(unknownIDs) > 0 {
md := wk.MonitoringMetadata(ctx, unknownIDs)
j.AddMetricShortIDs(md)
}
metricsStore.contributeMetrics
(sdks/go/pkg/beam/runners/prism/internal/jobservices/metrics.go) stores only
one payload per unresolved short ID in unprocessedPayloads [2]map[string][]byte:
m.unprocessedPayloads[d][short] = payload
wk.MonitoringMetadata is a gRPC call to the worker and runs without m.mu
held. When two or more concurrent bundles of the same stage finish before
AddMetricShortIDs runs, each later bundle overwrites
m.unprocessedPayloads[committed][short], and AddShortIDs replays only the last
bundle's payload. Data output is intact, but GetJobMetrics under-counts the
metric.
Fix:
Change unprocessedPayloads to [2]map[string][][]byte. Append committed
payloads for unresolved short IDs, replace tentative payloads with the latest
snapshot, replay buffered payloads in arrival order inside AddShortIDs, and
delete only resolved short IDs so other pending short IDs are preserved. Add
unit and race tests in metrics_test.go.
Reproduction (Python SDK on PrismRunner):
```python
import threading
import apache_beam as beam
from apache_beam.metrics.metric import Metrics
from apache_beam.options.pipeline_options import PipelineOptions
from apache_beam.transforms.trigger import AccumulationMode, AfterCount,
Repeatedly
from apache_beam.transforms.window import GlobalWindows
BARRIER = threading.Barrier(4)
class ConcurrentCounterDoFn(beam.DoFn):
def __init__(self):
self.counter = Metrics.counter("ns", "bundles")
def process(self, element):
BARRIER.wait(timeout=10)
self.counter.inc(1)
yield element
if __name__ == "__main__":
options = PipelineOptions(["--runner=PrismRunner",
"--environment_type=LOOPBACK"])
p = beam.Pipeline(options=options)
_ = (
p
| beam.Create([("k1", 1), ("k2", 1), ("k3", 1), ("k4", 1)])
| beam.WindowInto(
GlobalWindows(),
trigger=Repeatedly(AfterCount(1)),
accumulation_mode=AccumulationMode.DISCARDING,
)
| beam.GroupByKey()
| beam.ParDo(ConcurrentCounterDoFn())
)
res = p.run()
res.wait_until_finish()
counters = res.metrics().query()["counters"]
committed = counters[0].committed
print("committed:", committed)
assert committed == 4, f"expected 4 committed updates, got {committed}"
```
Before the fix: `committed` is `1` (or `< 4`), and the assertion fails.
After the fix: `committed: 4`.
<!--
Thank you for your contribution! Follow this checklist to help us
incorporate your contribution quickly and easily:
- [ ] If this contribution is large, please file an Apache [Individual
Contributor License Agreement](https://www.apache.org/licenses/icla.pdf).
- [ ] Replace these instructions with a description of your change and why
you have made it.
- [ ] Mention the appropriate issue in your description (for example:
`addresses #123` or `fixes #456`).
- [ ] Update `CHANGES.md` with noteworthy changes.
See [CONTRIBUTING.md]() for more details.
-->
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
