Discovery items type error failure

[jsetton]

Per the skill logs, several users discovery requests are failing due to the error below

error: discover failed with error:items.forEach is not a function
{
    "stack": "TypeError: items.forEach is not a function\n    at Promise.all.then (/var/task/alexa/v3/discovery.js:45:13)\n    at <anonymous>\n    at process._tickDomainCallback (internal/process/next_tick.js:228:7)"
}

After enabling debug mode, the issue seems to be related to the items data value being a string because the request call seems to only have read/received a partial response preventing the expected JSON object from being parsed properly.

The example based on the reference below isn't that considerably big in size, although it's just partial data, and the Lambda execution time for that request was about 5 seconds.

Lambda request log reference

[jsetton]

@digitaldan I moved this one into its own issue. I am still seeing this error being generated a couple times every hour.

Anyway, I did some research and I found a possible solution including Connection: keep-alive to the request headers.

I did some testing and the request module seems to defaults to Connection: close, if not specified, which would explain why the network connection could get prematurely closed.

Do you see any reason for not making this change?

Direct Connection

REQUEST onRequestResponse http://localhost:8080/rest/items?fields=editable%2CgroupNames%2CgroupType%2Cname%2Clabel%2Cmetadata%2Cstate%2CstateDescription%2Ctags%2Ctype&metadata=alexa%2Cchannel%2Csynonyms 200 { 
  connection: 'close',
  'content-type': 'application/json',
  server: 'Jetty(9.4.18.v20190429)'
}

Behind Reverse Proxy Server (nginx)

REQUEST onRequestResponse https://myopenhab.org/rest/items?fields=editable%2CgroupNames%2CgroupType%2Cname%2Clabel%2Cmetadata%2Cstate%2CstateDescription%2Ctags%2Ctype&metadata=alexa%2Cchannel%2Csynonyms 200 {
  server: 'nginx',
  date: 'Mon, 19 Aug 2019 16:41:07 GMT',
  'content-type': 'application/json',
  'transfer-encoding': 'chunked',
  connection: 'close'
}

[digitaldan]

If that works i don't see a problem with it. I'm not sure why it works :-) afaik 'keep-alive' is used so the underlying TCP connection can be used multiple times by different requests and save having to create and tear down a TCP connection over and over for each request, browsers would be painfully slow without it.

If the request is still in flight, and there is a pause in the data stream, i guess it's possible something like a nginx proxy could terminate it if the keep-alive is not set, so maybe thats what it is fixing?

[jsetton]

If the request is still in flight, and there is a pause in the data stream, i guess it's possible something like a nginx proxy could terminate it if the keep-alive is not set, so maybe thats what it is fixing?

My understanding is that when sending keep-alive to the server, it will not close the connection after responding to the initial request, unless the client requests it. While the connection remains open, it can be reused and will timeout after a specific time when idling. On top of this, I believe keep-alive is enabled by default in HTTP/1.1.

I think we should enable it/deploy and see if we still get that error. Do you see any reason why we shouldn't?

[digitaldan]

I am totally good with it.

[jsetton]

So the connection keep-alive change I made didn't resolve this issue.

After doing more research, I believe the issue is related to the fact that the myopenhab cloud connector nginx server is buffering every request/response before passing the data on.

That buffer has a limited size and can become full at time when there are a lot simultaneous discovery requests in a short amount of time. My understanding is that when the buffer gets full, some of the data chunks may be dropped in flight causing partial response to be sent.

Going over the nginx configuration committed in the cloud service github repository, I can see that the proxy_buffering setting is enabled by default, since omitted. Is there a reason why we want to have it enabled? It seems that disabling would potentially resolve our issue.

proxy_buffering off;

Also, we could enable gzip for application/json request type to decrease the response size being sent back to the Lambda function, which would also decrease the response time I believe.

gzip on;
gzip_comp_level 6;
gzip_types application/json;

@digitaldan What do you think?

[digitaldan]

I'll take a look at our nginx config, i'm not 100% sure if we set that for a reason or it was copied from some default nginx proxy config example.

[jsetton]

So since we now have the request timeout logic deployed, we get the bearer token for the request that timed out in the logs.

Based on my testing on a few of them, it seems that the request times out after 60 seconds when the myopenhab cloud connector nginx server times out with a "504 Gateway Time-out" error. Anyway, I was under the impression the cloud connector can track remote clients that are offline and directly respond to external requests with an offline message response.

The reason I am bringing this up is because each of the failed requests is wasting 8 seconds of compute time. There are about 35 timed out requests across the regions per hour so far. Maybe there is something that can be easily done on the cloud connector side.

[jsetton]

According to this documentation, if Cache-Control: no-cache is specified in the request headers, nginx will not cache responses.

Additionally, after doing some test on one of the affected users, I noticed that the proper JSON object was returned on subsequent calls. So we could introduce retry logic at that level if the items data isn't valid.