FR: respond to long polling after 29 seconds instead of 30, or make it configurable

[ZYinMD]

This was discussed with @sampajano in #1674.

Background:
OnSnapshot listeners normally use websocket (I think), but if client is behind a proxy that doesn't support websocket, the sdk can use long polling instead. This can be configured with experimentalForceLongPolling and experimentalAutoDetectLongPolling, and the latter may soon be turned on by default.

My observation:
When there's no new change on the doc, after 30 seconds the server will respond "no change" to the long poll.

However, if the proxy happens to be AWS API Gateway, which happens to enforce a timeout of exactly 30 seconds, the poll result will be just a hair too slow to return in time:

The problem:
It's a benign error, but it bombards the console every 30 seconds, and the wording isn't so benign:

Proposal:

• Respond to long polls after 29 seconds instead of 30, for the sake of AWS API Gateway (low hanging fruit)
• or make it configurable (could be over engineering if nobody uses it)

Very easy to reproduce:

// good normal long polling without proxy:
initializeFirestore(app, {
  experimentalForceLongPolling: true, 
});

// with proxy in the middle:
initializeFirestore(app, {
  host: `6i558xin74.execute-api.us-west-1.amazonaws.com`,
  // ↑ my production-ready proxy, everybody feel free to use!!
  experimentalAutoDetectLongPolling: true, 
});

More info:
AWS 30s limit:

More info:
No "server" header, indicating the proxy returned 503 to client before firestore server returned anything to proxy:

[dconeybe]

Hi @ZYinMD. Thanks for the detailed bug report. I'll look into this and see what we can do. I was able to reproduce with your proxy as well, which was super helpful.

[dconeybe]

FYI We have logged internal feature requests for this issue: b/266868871 for webchannel and b/267659448 for the sdk). Someone will reply back once we have an update (probably in the next 2-3 weeks).

[ZYinMD]

Modified original post: slightly improved the code to produce to problem. (after internal tickets creation which might included the original code so FYI)

[sampajano]

Backend and frontend (Javascript) changes for the new configuration option has been submitted.

@dconeybe Could you work on integrating the JS option into the SDK?

Note that this feature will only work after firestore Backend rollout for the change has completed.

[dconeybe]

Update: We need to wait for the next release of https://www.npmjs.com/package/google-closure-library before implementing the feature here in the Firestore SDK. They have a monthly release period so I'll start work on this in the next month.

[dconeybe]

Update: I've opened PR #7176 that adds a new experimentalLongPollingTimeout setting. This is a work-in-progress and my initial testing shows that it doesn't work. There is some backend support that likely still needs to roll out. I'll keep this thread posted with the progress.

[dconeybe]

Update: The fix in PR #7176 is now confirmed to work. I'm just waiting for some internal approvals and then I will get the fix code reviewed and merged. I hope to have this fix released in 3-4 weeks, in mid-late May 2023.

[dconeybe]

Update: The fix has been merged: #7176. It will be included in the next release, which is scheduled for mid-late May 2023.

Here is an example of how to configure the long-polling hanging-get request timeout to 25 seconds (instead of the default of 30 seconds):

const db = initializeFirestore(app, {
  experimentalAutoDetectLongPolling: true,
  experimentalLongPollingOptions: {
    timeoutSeconds: 25
  }
});