Kenneth Knowles created BEAM-6997:
-------------------------------------
Summary: LTS backport: No messages delivered after a while with
PubsubIO
Key: BEAM-6997
URL: https://issues.apache.org/jira/browse/BEAM-6997
Project: Beam
Issue Type: Bug
Components: io-java-gcp, runner-flink
Affects Versions: 2.4.0
Reporter: Hrish
Assignee: Maximilian Michels
Fix For: 2.11.0
I am running the following Beam pipeline code locally, with the FlinkRunner.
PubsubIO is used to read messages from a topic. I have a separate thread that
publishes messages to the topic at regular intervals (every 30 seconds) and
also sets the "ts" attribute which is used later to derive the event time.
Custom transform to convert to KV pair -
{code:java}
private static class PubSubMessageGrouper extends DoFn<PubsubMessage,
KV<String, PubsubMessage>> {
@ProcessElement
public void processElement(ProcessContext c) {
PubsubMessage element = c.element();
KV<String, PubsubMessage> kv = KV.of(element.getAttribute("key"),
element);
c.output(kv);
}
}
{code}
Note that "key" is a key set in the message attributes earlier in the publisher
thread. The intent is to group the messages downstream by this key.
Pipeline code -
{code:java}
PCollection<PubsubMessage> pubsubColl = p
.apply(PubsubIO.readMessagesWithAttributes()
.withTimestampAttribute("ts")
.fromTopic("projects/" + projectName + "/topics/beamtest")
);
PCollection<KV<String, PubsubMessage>> idfied =
pubsubColl.apply(ParDo.of(new PubSubMessageGrouper()));
PCollection<KV<String, PubsubMessage>> windowed = idfied
.apply(Window.<KV<String,
PubsubMessage>>into(FixedWindows.of(Duration.standardSeconds(15)))
.triggering(
Repeatedly.forever(
AfterWatermark.pastEndOfWindow()
)
)
.withAllowedLateness(Duration.standardSeconds(15))
.discardingFiredPanes());
PCollection<KV<String, Iterable<PubsubMessage>>> grouped =
windowed.apply(GroupByKey.create());
grouped.apply(ParDo.of(new KVPrinter()));
{code}
The transforms are not chained for ease of reading. The KVPrinter transform in
the end is just to print out the messages received from the group by, which
will be subsequently replaced by actual code once I get this running. When I
run this, I don't find the trigger executing for quite some time (a couple of
minutes or longer). When it finally triggers, I see that some of the messages
are not received (in the final step), not matter how long I keep it running.
The Pubsub statistics in my GCP/Stackdriver dashboard show that there is a
backlog of undelivered messages.
Is this due to the internal watermark that PubsubIO uses? My intention here is
to make sure that all messages are processed in the groupby, including late
ones within the allowed lateness window.
Note that if I remove the GroupByKey, and just print the messages after the
windowing, I can see all the messages.
--------
An update: I switched the .fromTopic to a .fromSubscription, with a
pre-created subscription, and things have started working, as in, I can see
messages being delivered now. The only difference I can see is that the
automatically created subscription from within PubsubIO has an ack deadline of
60 seconds, whereas I created mine with 600 seconds (the max allowed by pubsub).
However, there is another issue now - messages don't seem to be getting acked.
I keep getting redeliveries of old messages, and the Stackdriver metrics show
that the num_undelivered_messages (unacked messages count) keeps on increasing.
The documentation says that messages will be acked as soon as a GroupByKey or
another ParDo happens, and that's happening since I am seeing the windowed
groupby-s, but acks don't seem to be happening.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
