zentol commented on a change in pull request #6850: [FLINK-10252] Handle
oversized metric messges
URL: https://github.com/apache/flink/pull/6850#discussion_r230297487
##########
File path:
flink-runtime/src/main/java/org/apache/flink/runtime/metrics/dump/MetricDumpSerialization.java
##########
@@ -124,55 +160,135 @@ public MetricSerializationResult serialize(
Map<Counter, Tuple2<QueryScopeInfo, String>> counters,
Map<Gauge<?>, Tuple2<QueryScopeInfo, String>> gauges,
Map<Histogram, Tuple2<QueryScopeInfo, String>>
histograms,
- Map<Meter, Tuple2<QueryScopeInfo, String>> meters) {
+ Map<Meter, Tuple2<QueryScopeInfo, String>> meters,
+ long maximumFramesize) {
+
+ boolean markUnserializedMetrics = false;
- buffer.clear();
+ Map<Counter, Tuple2<QueryScopeInfo, String>>
unserializedCounters = new HashMap<>();
+ Map<Gauge<?>, Tuple2<QueryScopeInfo, String>>
unserializedGauges = new HashMap<>();
+ Map<Histogram, Tuple2<QueryScopeInfo, String>>
unserializedHistograms = new HashMap<>();
+ Map<Meter, Tuple2<QueryScopeInfo, String>>
unserializedMeters = new HashMap<>();
+ countersBuffer.clear();
int numCounters = 0;
for (Map.Entry<Counter, Tuple2<QueryScopeInfo, String>>
entry : counters.entrySet()) {
+ if (markUnserializedMetrics) {
+
unserializedCounters.put(entry.getKey(), entry.getValue());
+ continue;
+ }
+
try {
- serializeCounter(buffer,
entry.getValue().f0, entry.getValue().f1, entry.getKey());
+ serializeCounter(countersBuffer,
entry.getValue().f0, entry.getValue().f1, entry.getKey());
numCounters++;
+ if (countersBuffer.length() >
maximumFramesize) {
+ LOG.warn("The serialized
counter metric is larger than the maximum frame size, " +
+ " so maybe not all
metrics would be reported.");
+ markUnserializedMetrics = true;
+ //clear all, because we can not
revoke the latest metrics which caused overflow
+
unserializedCounters.put(entry.getKey(), entry.getValue());
+ countersBuffer.clear();
+ numCounters = 0;
+ }
} catch (Exception e) {
LOG.debug("Failed to serialize
counter.", e);
}
}
+ gaugesBuffer.clear();
int numGauges = 0;
for (Map.Entry<Gauge<?>, Tuple2<QueryScopeInfo,
String>> entry : gauges.entrySet()) {
+ if (markUnserializedMetrics) {
+ unserializedGauges.put(entry.getKey(),
entry.getValue());
+ continue;
+ }
+
try {
- serializeGauge(buffer,
entry.getValue().f0, entry.getValue().f1, entry.getKey());
+ serializeGauge(gaugesBuffer,
entry.getValue().f0, entry.getValue().f1, entry.getKey());
numGauges++;
+ if (gaugesBuffer.length() +
countersBuffer.length() > maximumFramesize) {
+ LOG.warn("The serialized gauge
metric is larger than the maximum frame size, " +
+ " so maybe not all
metrics would be reported.");
+ markUnserializedMetrics = true;
+
unserializedGauges.put(entry.getKey(), entry.getValue());
+ gaugesBuffer.clear();
+ numGauges = 0;
+ }
} catch (Exception e) {
LOG.debug("Failed to serialize gauge.",
e);
}
}
+ histogramsBuffer.clear();
int numHistograms = 0;
for (Map.Entry<Histogram, Tuple2<QueryScopeInfo,
String>> entry : histograms.entrySet()) {
+ if (markUnserializedMetrics) {
+
unserializedHistograms.put(entry.getKey(), entry.getValue());
+ continue;
+ }
+
try {
- serializeHistogram(buffer,
entry.getValue().f0, entry.getValue().f1, entry.getKey());
+ serializeHistogram(histogramsBuffer,
entry.getValue().f0, entry.getValue().f1, entry.getKey());
numHistograms++;
+ if (histogramsBuffer.length() +
gaugesBuffer.length() + countersBuffer.length() > maximumFramesize) {
+ LOG.warn("The serialized
histogram metric is larger than the maximum frame size, " +
+ " so maybe not all
metrics would be reported.");
+ markUnserializedMetrics = true;
+
unserializedHistograms.put(entry.getKey(), entry.getValue());
+ histogramsBuffer.clear();
+ numHistograms = 0;
+ }
} catch (Exception e) {
LOG.debug("Failed to serialize
histogram.", e);
}
}
+ metersBuffer.clear();
int numMeters = 0;
for (Map.Entry<Meter, Tuple2<QueryScopeInfo, String>>
entry : meters.entrySet()) {
+ if (markUnserializedMetrics) {
+ unserializedMeters.put(entry.getKey(),
entry.getValue());
+ continue;
+ }
+
try {
- serializeMeter(buffer,
entry.getValue().f0, entry.getValue().f1, entry.getKey());
+ serializeMeter(metersBuffer,
entry.getValue().f0, entry.getValue().f1, entry.getKey());
numMeters++;
+ if (metersBuffer.length() +
histogramsBuffer.length() + gaugesBuffer.length() +
Review comment:
Let's say that our max frameSize is 100 bytes. Let's also say that counters,
meters and gauges take up 1 byte, and histograms take up 98.
I the order that you're serializing metrics (counters -> gauges ->
histograms -> meters), counters, gauges and histograms will be serialized and
fit (1 + 1 + 98 <= 100). Then the meters come along with their additional 1
byte and are now dropped from the report.
Given that the only known case of this problem arising is a ridiculous
amount of latency histograms they should be excluded first, instead of meters.
As it stands histograms are the least important yet most expensive metrics, so
let's drop those first instead of rather essential metrics like
`numRecordsInPerSecond`.
Now, this does _not_ mean that you now should change the order again (like
you had in a previous version), but that the MQS should first try to drop
histograms to fit the report into the dump.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
