Google-cloud-node: subscription.on message wasn't received of messagActually unrelated failures, but we should bump this anyway.es after 10 mins or more.

[CoNETProject]

• OS: Debian
• Node.js version: 8.4.0
• npm version: 5.4.0
• google-cloud-node version: 0.56.0
• @google-cloud/pubsub version: 0.14.0

Steps to reproduce

const pubsub = require ( '@google-cloud/pubsub' )( connectOption )

const subscription = pubsub.topic ( topicName ).subscription ( subScription )

subscription.on ('message', message => {
    message.ack()
    ....
})

subscription.once('error', err => {
    ....
})

I received all un-received messages when I restart it.
I checked the connectionPool object when it have not received message:
connectionPool.isPaused [false]
connectionPool.isOpen [true]
connectionPool.connections [5]

Thank you.

[callmehiphop]

Thanks for reporting! We've had multiple reports of similar behavior and we're currently investigating this issue. In the meantime you might want to use pubsub v0.13.x as a temporary workaround.

[CoNETProject]

Thank you.
I did a temp code to use the pubsub v0.14.0 like this:

const pubsub = require ( '@google-cloud/pubsub' )( connectOption )

const listenSub = () => {
    let subscription = pubsub.topic ( topicName ).subscription ( subScription )
    subscription.on ( 'message', message => {
        message.ack()
        ....
   })

    subscription.once('error', err => {
        ....
    })

    //        each 10 mins restart
    const tt = setTimeout (() => {
    return subscription.close().then(() => {
        listenPubSubMaster = null
        return listenSub ()
    }, 600000 )
}

listenSub()

[caseyduquettesc]

We see a similar behavior, although it takes a few hours. Eventually the subscription stops processing new messages and we have to restart the subscription. This only started after upgrading to 0.14.

[mscottx88]

I too am seeing the same problem. We set up a forever running process on GKE and it seems to go unresponsive after some time. I came to the same conclusion of putting a restarter.

Let's hope this gets resolved soon!

[callmehiphop]

This is definitely a result of moving towards the new streaming API, I'm working on resolving this issue, but in the mean time I suggest downgrading to 0.13. I'll update this issue once it's been resolved.

[pongad]

@callmehiphop FYI, we resolved this in Go by adding gRPC keepalive. I'm not sure how the code should look like in Node though. https://code-review.googlesource.com/#/c/gocloud/+/17070/

[callmehiphop]

@pongad thanks for the update! I believe we just have to set the grpc.keepalive_time_ms option.

[callmehiphop]

@QTGate @mscottx88 @caseyduquettesc I've opened a PR (#2627) which I believe will fix the issue you are seeing, would any of you mind testing it out? Thanks!

[mscottx88]

I'm willing to give it a go, for sure!

On Mon, Sep 25, 2017, 7:17 AM Dave Gramlich [e-mail]
wrote:

@QTGate https://github.com/qtgate @mscottx88
https://github.com/mscottx88 @caseyduquettesc
https://github.com/caseyduquettesc I've opened a PR (#2627
https://github.com/GoogleCloudPlatform/google-cloud-node/pull/2627)
which I believe will fix the issue you are seeing, would any of you mind
testing it out? Thanks!

—
You are receiving this because you were mentioned.

Reply to this email directly, view it on GitHub
https://github.com/GoogleCloudPlatform/google-cloud-node/issues/2598#issuecomment-331895506,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AT6MMUjOHHVuq1ZzzlIViODrOF50HIecks5sl7X8gaJpZM4PQKSI
.

[mscottx88]

@callmehiphop I applied the two fixed files to my deployment but now I am seeing duplication of message delivery. That is, I am seeing the same message being pulled more than one time.

update Also, it appears that the connection has been lost or something as it is sitting idle again even with messages being published to it.

[callmehiphop]

@jganetsk I'm noticing that messages that are not acked before a stream is closed will be redelivered when the next stream is opened regardless of whether or not the ack deadline was hit. Is that the expected behavior? If so I'm assuming we need to filter these messages out.

[jganetsk]

@callmehiphop I don't think you should be filtering any messages out. In general, it is possible to receive duplicates, and it is possible that they come before the ack deadline was hit. If this happens consistently, that's a problem that we should look at. How much time passes between delivery and redelivery? Does it correspond to the default ack deadline?

[jganetsk]

We will take a look at this, thanks for reporting it.

[callmehiphop]

@jganetsk as far as I can tell it only happens when closing and re-opening a stream. I'm putting together an elaborate test and I'll give you the details of my findings once I have some to share.

[callmehiphop]

There appears to be 2 different scenarios in which messages will be redelivered before they have been acked.

Streams being closed.

In the event of a stream being closed it appears that all un-acked messages will be redelivered to a different stream regardless of the current ack deadline.

• Test 1: Set the initial streamAckDeadlineSeconds to 60 (1 minute) and close a stream after 30 seconds.
  Result: All messages were redelivered despite not hitting the deadline.
• Test 2: Modify the deadline just before closing the stream.
  Result: All messages were redelivered even though the deadline was successfully modified.

Race Condition

I believe we occasionally hit a race condition when attempting to modify an existing ack deadline. Currently we calculate the ack deadline using the 99th percentile of ack times (p99). We then calculate a delay to determine when we should modify the ack deadline (Math.random() * n99 * 0.9).

If our p99 value is 10, then it is possible we will not modify the ack deadline until roughly 1 second before it hits. The redelivery cases I have observed occurred when the delay was within 1-2 seconds of the p99 value.

@lukesneeringer have you seen anything similar to what I'm describing above in the Python client? If this is indeed the reason why a redelivery occurs, is it reasonable to calculate the sleep time at a lower percent?

[mscottx88]

Here is my solution to the race condition you describe. A simple activeMessages tracking mechanism, combined with async shutdown that is event-driven.

'use strict';

const config = require('@google-cloud/rcloadenv');
const cp = require('child_process');
const fs = require('fs-extra');
const logger = require('winston');
const pubsub = require('./providers/pubsub');
const uuid = require('uuid/v4');

const { name: configurationStorageName } = require('../package.json');

/**
 * Listen for and process incoming messages.
 *
 * @returns {Function}
 * An async function that can be called to stop listening for messages.
 */
const listen = () => {
  let activeMessages = 0;
  logger.info(`Subscription started at ${new Date().toISOString()}.`);

  const subscription = pubsub
    .get()
    .topic(process.env.SUBSCRIBE_TOPIC_NAME)
    .subscription(process.env.SUBSCRIBE_SUBSCRIPTION_NAME, {
      flowControl: {
        maxMessages: +process.env.MAX_CONCURRENT_MESSAGES
      }
    });

  subscription.on('error', (error) => {
    logger.error(error);
  });

  subscription.on('message', (message) => {
    const { email, pwd, uid } = message.attributes;
    logger.info(`Message received at ${new Date().toISOString()} for uuid ${uid}.`);

    const sessionId = uuid();
    const downloadPath = `${__dirname}/../downloads/${sessionId}`;
    const profilePath = `${__dirname}/../profiles/${sessionId}`;

    const args = [
      uid,
      email,
      pwd,
      sessionId,
      downloadPath,
      profilePath
    ];

    const options = {
      cwd: process.cwd(),
      env: process.env,
      execArgv: [],
      silent: true
    };

    logger.info(`Starting session ${sessionId}`);
    const downloader = cp.fork(`${__dirname}/download.js`, args, options);
    activeMessages++;

    downloader.stdout.on('data', (data) => {
      const text = new Buffer(data).toString('utf8');
      logger.info(text);
    });

    downloader.on('exit', async (code) => {
      logger.info(`Session ${sessionId} completed with exit code ${code}.`);

      try {
        await Promise.all([fs.remove(downloadPath), fs.remove(profilePath)]);
      } catch (error) {
        logger.error(`Could not cleanup path(s) in session ${sessionId} because of ${error}.`);
      }

      try {
        message.ack();
      } finally {
        activeMessages--;
      }
      if (activeMessages === 0) {
        subscription.emit('importer-idle');
      }
    });
  });

  return () => {
    return new Promise((resolve, reject) => {
      // stop the flow of incoming messages
      subscription.removeAllListeners('message');

      const shutdown = async () => {
        try {
          await subscription.close();
          logger.info(`Subscription ended at ${new Date().toISOString()}.`);
          resolve();
        } catch (error) {
          reject(error);
        }
      };

      if (activeMessages !== 0) {
        subscription.on('importer-idle', shutdown);
      } else {
        shutdown();
      }
    });
  };
};

/**
 * Run the process.
 *
 * @returns {void}
 */
const run = async () => {
  logger.level = process.env.LOG_LEVEL;

  try {
    await config.getAndApply(configurationStorageName);
  } catch (error) {
    logger.error(error);
    throw error;
  }

  let unlisten = listen();

  setInterval(async () => {
    try {
      await unlisten();
    } catch (error) {
      logger.error(error);
    } finally {
      unlisten = listen();
    }
  }, +process.env.RESTART_INTERVAL);
};

module.exports = {
  run
};

[CoNETProject]

• OS: Debian
• Node.js version: 8.7.0
• npm version: 5.5.1
• @google-cloud/pubsub version: 0.14.5

Steps to reproduce

const pubsub = require ( '@google-cloud/pubsub' )( connectOption )

const listenSub = () => {
    let subscription = pubsub.topic ( topicName ).subscription ( subScription )
    subscription.on ( 'message', message => {
        message.ack()
        ....
   })

    //        each 10 mins restart
    const tt = setTimeout (() => {
    return subscription.close().then(() => {
        listenPubSubMaster = null
        return listenSub ()
    }, 600000 )
}

listenSub()

I did make new subscription each 10 mins.
But looks the @google-cloud/pubsub did not release memory when pubsub exit with subscription.close().

[ShahNewazKhan]

• OS: Debian
• Node.js version: 6.11.1
• npm version: 3.10.10
• version: 0.14.5

I am also seeing the memory increase and not being released after implementing the subscriber refresh.

[stephenplusplus]

@callmehiphop will this be addressed by #2716 or another issue/PR you've already worked on?

[CoNETProject]

No I am not. Even the new v0.14.6. Also have memory problem. I plan do subscription.on gCloud API's by my self.

[callmehiphop]

will this be addressed by #2716

I believe it will!

[lc-chrisbarton]

I am using PubSub node.js client version 0.14.5 and I sometimes have CloudFunctions that when triggered are no longer able to add messages to a topic or pull messages from a subscription. The fix is to redeploy the CloudFunction which mostly works but not always. The issue is intermittent.

[rhodgkins]

Apologies for asking the obvious...!

But is this now fixed as #2716 has been released with 0.14.8?

Cheers!

[callmehiphop]

@rhodgkins which parts? I think there are a number of issues being described here.

[rhodgkins]

@callmehiphop the issue around no messages being received on a subscription after a period of time.

[callmehiphop]

@rhodgkins from what I understand the only environment in which this occurs is GKE, but I also know a fix was deployed on the GKE side of things that should have resolved it. I haven't seen or heard of any reports verifying this though.

[stephenplusplus]

This issue was moved to googleapis/nodejs-pubsub#10

[mchavarriae]

Recently we start experimenting the problem using version 0.13.0, Does anybody has this problem?