Subscriptions-transport-ws: Connection timeout strategy causing disconnections

[dotansimha]

@mistic

Since the PR that adds the following:
https://github.com/apollographql/subscriptions-transport-ws/blob/v0.8.3/src/client.ts#L407-L416

The default connection timeout starts with 1000ms which isn't enough for low performance servers (for example: free plan in heroku or free plan in now).
This causing the client to disconnect if it took more than 1000ms to perform the connection.

For example, with my remote server, I expect the client to connect in about 1500ms or more - in this case I'm seeing one or more disconnections every time I recreate my socket.

The 1000ms should be configurable, and I think it should be more than 1000ms by default.

apollo-client: 1.9.3
transport: 0.8.3

[pandemosth]

I'm seeing this at the moment too (also on heroku). To add to the issue, appears that the reconnect logic is broken if the max connect timeout fires.

In my scenario the lazy flag is set, so first subscription is triggering the websocket connection attempt. If this times-out and is reattempted I'm seeing two different behaviours after the connection is opened:

1. No subscription has been established (on Android real device)
2. Two instances of the subscription that both then fire on new data from the server (on iOS Simulator and real device)

Haven't got any further on why this is happening but will continue to investigate.

And this is react-native with following versions:

apollo-client: 1.9.3
react-apollo: 1.4.16

[pandemosth]

OK, so I've managed to fix both issues above, changes here for comment and advice on whether this should be a PR.

@dotansimha I can create a separate issue if I'm hijacking this one..?

https://github.com/apollographql/subscriptions-transport-ws/compare/master...pandemosth:issue295

Explanation follows:-

Issue 1 - No subscription is made on reconnect, after initial connect times out (only seen on Android).
This is a strange one, the WebSocket onOpen callback is being called but the readyState is still 0 (connecting). This causes the following (apologies for verbose explanation):

• sendMessage is called from onOpen, then in sendMessageRaw status is CONNECTING so message is pushed onto unsent queue
• then next line in onOpen calls flushUnsentMessagesQueue. The subscription message and the init message in the unsent queue are attempted again, but since status is CONNECTING they are not sent and are added back to the unsent queue. But, in flushUnsentMessagesQueue the queue is then set to an empty array, so the unsent messages are discarded.
• onOpen is called again, this time with status 1, the websocket is connected but no graphql subscription is established since the unsent queue is empty.

This is probably an issue with React Native / WebSocket on Android. Seems odd that onOpen would be called with reedyState still 0. But, easy enough to guard against that in code which is what I've done.

Issue 2 - Duplicate subscription made on reconnect, after initial connect timeout (both platforms).
After fixing issue 1, this one was observed consistently on both platforms. The max connect timeout calls close which calls tryReconnect. In tryReconnect all current operations are added onto the unsent queue, which is causing the duplication. The simple fix here is to set this.reconnecting = true in the max connect timeout callback in order to bypass this step.

I couldn't see any tests in the project related to timeouts so haven't updated tests.

Btw, Charles proxy with throttling (+1000ms latency) was used to reproduce these issues and verify.

[perrosnk]

Any updates on this? I am having the same issues

[ilijaNL]

@pandemosth can confirm this on android, sometimes a dubbel subscription is made, or none at all

[ash0080]

Same issues, no subscription or duplicated subscriptions, any news here?

[Amareis]

377 solve this

[nwronski]

I was also experiencing Issue 2 described by @pandemosth and can confirm his changes fixed the issue. I observed that multiple copies of the initial operation(s) were sent to the server whenever a websocket connection took longer than 1000ms to connect.

[NeoPhi]

@pandemosth or @nwronski Please open a PR with the changes. The existing logic definitely sounds buggy.

[jpgarcia]

@NeoPhi If you want I can create a PR with the changes @pandemosth suggested against master.

I've just forked the repo and pushed https://github.com/underscopeio/subscriptions-transport-ws/commit/15c0d70d1c4e20031a7635270301eb84dbdd864f which is pretty much the same as @pandemosth except the src/server.ts which AFAIK is already updated on master.

I've tested the fix in a react-native project and it seems to be working fine now! (thank you @pandemosth !)

Once published the new version we should also create an issue in apollo-link-ws to upgrade the subscriptions-transport-ws dependency

[mxstbr]

@jpgarcia PR would be much appreciated!

[jpgarcia]

@mxstbr done!

[NeoPhi]

Included in: https://github.com/apollographql/subscriptions-transport-ws/releases/tag/v0.9.12

[jedwards1211]

@NeoPhi not everything in OP is fixed, it still appears we don't have a good way to raise the initial connect timeout do we?

I bet even just an expensive initial page load can overwhelm the 1 second initial connect timeout, even if the server is performing just fine.

[jedwards1211]

Also, is applying the backoff to the connection timeout really the best way to go about it? Seems to me the backoff should apply to the time waited before attempting a reconnect, instead of the time waited for the connection to succeed.