[
https://issues.apache.org/jira/browse/GOBBLIN-1934?focusedWorklogId=886870&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-886870
]
ASF GitHub Bot logged work on GOBBLIN-1934:
-------------------------------------------
Author: ASF GitHub Bot
Created on: 24/Oct/23 05:59
Start Date: 24/Oct/23 05:59
Worklog Time Spent: 10m
Work Description: phet commented on code in PR #3805:
URL: https://github.com/apache/gobblin/pull/3805#discussion_r1369656381
##########
gobblin-runtime/src/main/java/org/apache/gobblin/runtime/metrics/RuntimeMetrics.java:
##########
@@ -28,6 +28,7 @@ public class RuntimeMetrics {
// Metric names
public static final String GOBBLIN_KAFKA_HIGH_LEVEL_CONSUMER_MESSAGES_READ =
"gobblin.kafka.highLevelConsumer.messagesRead";
+ public static final String GOBBLIN_KAFKA_HIGH_LEVEL_CONSUMER_QUEUE_SIZE =
"gobblin.kafka.highLevelConsumer.queueSize";
Review Comment:
nit: seems to be used as a prefix
##########
gobblin-runtime/src/main/java/org/apache/gobblin/runtime/kafka/HighLevelConsumer.java:
##########
@@ -197,6 +201,22 @@ protected void shutdownMetrics() throws IOException {
*/
protected void createMetrics() {
this.messagesRead =
this.metricContext.counter(RuntimeMetrics.GOBBLIN_KAFKA_HIGH_LEVEL_CONSUMER_MESSAGES_READ);
+ this.queueSizeGauges = new ContextAwareGauge[numThreads];
+ this.queueSizeGaugeValues = new AtomicIntegerArray(numThreads);
+ for (int i=0; i < numThreads; i++) {
+ int finalI = i;
+ this.queueSizeGauges[i] = this.metricContext.newContextAwareGauge(
+ RuntimeMetrics.GOBBLIN_KAFKA_HIGH_LEVEL_CONSUMER_QUEUE_SIZE + "-" +
i,
+ () -> queueSizeGaugeValues.get(finalI));
Review Comment:
w/o a comment explaining the reason, I'm left to wonder: why couldn't this
be written:
```
() -> queues[finalI].size()
```
? that would dispense w/ the need for `queueSizeGaugeValues` and with it
the potential perf hit of updating that array upon every enqueing.
and on that topic... why is `updateQueueSizes` only recalculated at enqueue
time, rather than at dequeue time as well?
##########
gobblin-runtime/src/main/java/org/apache/gobblin/runtime/kafka/HighLevelConsumer.java:
##########
@@ -197,6 +201,22 @@ protected void shutdownMetrics() throws IOException {
*/
protected void createMetrics() {
this.messagesRead =
this.metricContext.counter(RuntimeMetrics.GOBBLIN_KAFKA_HIGH_LEVEL_CONSUMER_MESSAGES_READ);
+ this.queueSizeGauges = new ContextAwareGauge[numThreads];
+ this.queueSizeGaugeValues = new AtomicIntegerArray(numThreads);
+ for (int i=0; i < numThreads; i++) {
+ int finalI = i;
Review Comment:
I see what you're doing... but it may not immediately register to
everyone... especially since `finalI` is not `final int`. please add a comment
calling attention to variable capture by the closure
##########
gobblin-runtime/src/main/java/org/apache/gobblin/runtime/kafka/HighLevelConsumer.java:
##########
@@ -197,6 +201,22 @@ protected void shutdownMetrics() throws IOException {
*/
protected void createMetrics() {
this.messagesRead =
this.metricContext.counter(RuntimeMetrics.GOBBLIN_KAFKA_HIGH_LEVEL_CONSUMER_MESSAGES_READ);
+ this.queueSizeGauges = new ContextAwareGauge[numThreads];
+ this.queueSizeGaugeValues = new AtomicIntegerArray(numThreads);
+ for (int i=0; i < numThreads; i++) {
+ int finalI = i;
+ this.queueSizeGauges[i] = this.metricContext.newContextAwareGauge(
+ RuntimeMetrics.GOBBLIN_KAFKA_HIGH_LEVEL_CONSUMER_QUEUE_SIZE + "-" +
i,
+ () -> queueSizeGaugeValues.get(finalI));
+ }
+ }
+
+ /**
+ * Update atomic integer corresponding to queue size
+ * @param queueIndex used to index the queue list
+ */
+ private void updateQueueSizes(int queueIndex) {
+ this.queueSizeGaugeValues.set(queueIndex, queues[queueIndex].size());
Review Comment:
since `AtomicIntegerArray::set` is not actually an atomic operation (of the
kind `getAndIncrement` or even `getAndSet`), do we truly need such an instance?
##########
gobblin-runtime/src/main/java/org/apache/gobblin/runtime/kafka/HighLevelConsumer.java:
##########
@@ -197,6 +201,22 @@ protected void shutdownMetrics() throws IOException {
*/
protected void createMetrics() {
this.messagesRead =
this.metricContext.counter(RuntimeMetrics.GOBBLIN_KAFKA_HIGH_LEVEL_CONSUMER_MESSAGES_READ);
+ this.queueSizeGauges = new ContextAwareGauge[numThreads];
+ this.queueSizeGaugeValues = new AtomicIntegerArray(numThreads);
+ for (int i=0; i < numThreads; i++) {
+ int finalI = i;
+ this.queueSizeGauges[i] = this.metricContext.newContextAwareGauge(
+ RuntimeMetrics.GOBBLIN_KAFKA_HIGH_LEVEL_CONSUMER_QUEUE_SIZE + "-" +
i,
+ () -> queueSizeGaugeValues.get(finalI));
+ }
+ }
+
+ /**
+ * Update atomic integer corresponding to queue size
+ * @param queueIndex used to index the queue list
+ */
+ private void updateQueueSizes(int queueIndex) {
Review Comment:
nit: there's only one index updated, so `Sizes` should be singular
##########
gobblin-runtime/src/main/java/org/apache/gobblin/runtime/kafka/HighLevelConsumer.java:
##########
@@ -237,13 +257,16 @@ protected void startUp() {
private void consume() {
try {
Iterator<KafkaConsumerRecord> itr = gobblinKafkaConsumerClient.consume();
+ // TODO: we may be committing too early and only want to commit after
process messages
if(!enableAutoCommit) {
commitOffsets();
}
while (itr.hasNext()) {
KafkaConsumerRecord record = itr.next();
int idx = record.getPartition() % numThreads;
queues[idx].put(record);
+ // Increment queue size metric
+ updateQueueSizes(idx);
Review Comment:
this is a pretty fundamental change to put into the fast path! we gotta be
certain that no critical use is adversely affected by perf regression.
I see `ConcurrentLinkedQueue::size` may not be `O(1)`, but I can't find a
statement on `LinkedBlockingQueue::size`...
(it's possible it IS `O(1)`, but an approximation or else exact, but at the
expense of coordination overhead.)
Issue Time Tracking
-------------------
Worklog Id: (was: 886870)
Time Spent: 20m (was: 10m)
> Monitor High Level Consumer queue size
> --------------------------------------
>
> Key: GOBBLIN-1934
> URL: https://issues.apache.org/jira/browse/GOBBLIN-1934
> Project: Apache Gobblin
> Issue Type: Improvement
> Components: gobblin-service
> Reporter: Urmi Mustafi
> Assignee: Abhishek Tiwari
> Priority: Major
> Time Spent: 20m
> Remaining Estimate: 0h
>
> We suspect that the HighLevelConsumer class may be dropping Kafka messages if
> the thread for a particular consumption queue is interrupted or stalled.
> Before changing how these queues are handled, we want to verify that blocked
> or dropped queues is the reason for missed messages by emitting metrics about
> each queue's size.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)