Google Pub/Sub 如何在 Pull 上设置读取超时

Question

I would like to set the read timeout of the pull request on a subscription. Right now the only options are to set returnImmediately=true or just wait until the pubsub returns, which seems to be 90 seconds if no messages is published.

I'm using the gcloud-node module to make calls to pubsub. It uses the request module under the hood to make the the gcloud api calls. I've updated my local copy of gcloud-node/lib/pubsub/subscription.js to set the request timeout to 30 seconds

this.request({
  method: 'POST',
  uri: ':pull',
  timeout: 30000,
  json: {
    returnImmediately: !!options.returnImmediately,
    maxMessages: options.maxResults
  }
}

When I do this, the behavior I see is the connection will timeout on the client side after 30 seconds, but pubsub still has the request open. If I have two clients pulling on the subscription and one of them timeout after 30 seconds, then a message is published to the topic, it is a 50/50 chance that the remaining listening client will retrieve the message.

Is there a way to tell pubsub to timeout pull connections after a certain amount of time?

UPDATE: I probably need to clarify my example a bit. I have two clients that connect at the same time and pull from the same subscription. The only difference between the two is that the first one is configured to timeout after 30 seconds. Since two clients are connected to the same subscription, pubsub will distribute the message load between the two of them. If I publish a message 45 seconds after both clients connect, there is a 50/50 chance that pubsub will deliver the message to the second client that has not timed out yet. If I send 10 messages instead of just one, the second client will receive a subset of the 10 messages. It looks like this is because my clients are in a long poll. If the client disconnects, the server has no idea and will try to send published messages on the response of the request that was made by the client that has timed out. From my tests, this is the behavior I've observed. What I would like to do is be able to send a timeout param in the pull request to tell subpub to send back a response after a 30000ms if no messages are published during that time. Reading over the API docs , this doesn't seem like an option.

Answer 1

Setting the request timeout as you have is the correct way to timeout the pull after 30 seconds. The existence of the canceled request might not be what is causing the other pull to not get the message immediately. If your second pull (that does not time out) manages to pull other messages that were published earlier, it won't necessarily wait for additional message that was published after the timeout to come in before completing. It only guarantees to not return more than maxMessages , not to return only once it has exactly maxMessages (if that many are available). Once your publish completes, some later pull will get the message, but there are no guarantees on exactly when that will occur.