Connection reset by peer (httplib2 issue)

[znorris]

I've been attempting to batch upload just over a million entities (split into groups of 100-500 at a time) into datastore but I keep running into a socket.error: [Errno 104] Connection reset by peer error. I've also been experiencing this when attempting to query and pull down all entities of a single kind.

Is there some throttling that should be happening that's not? Any suggestions on how to better utilize the gcloud lib?

[daspecster]

Hello @znorris!
Is the datastore client being run multi-threaded?

[znorris]

@daspecster No it's not.

[daspecster]

In that case I'm not exactly sure what is causing this. @dhermes or @tseaver might have some ideas.

If you retry after the socket error does it work? Could it possibly be a connectivity issue?

[dhermes]

Is there a long timeout before the socket error? Connection reset by peer usually means the handshake has broken down (often due to a long delay between requests). If the backend were unhappy with your requests, you'd likely get a status code with a rate-limit error.

[znorris]

@dhermes that's what I would have thought, some kind of msg telling me I'd hit a limit.

There are no long pauses between batch uploads so I assume it would not timeout there. Perhaps I'm hitting a limit on connection duration?

Export to Datastore stats:
Aprox. entities: 1.08 Million
Properties per entity: 2 (1 string, 1 boolean)
Aprox. size: 100 MB
Entities per put_multi(): 400
Upload duration: 3 hours 10 minutes
Local Bandwidth: 100 Mbps (upload & download)

It is taking much longer than expected to batch upload entities. The library makes it so simple to use I can't think of anything I could be doing differently on my end.

[dhermes]

@znorris Are you still having this issue? When you say "Upload duration: 3 hours 10 minutes", does that mean for a single request or for a batch of them?

[znorris]

@dhermes That time was for all batch operations to complete (all 1.08 Million entities).

[dhermes]

Gotcha. Any idea how long each request is taking? Can you share some code?

[dhermes]

Also reported in #2166 with logging. In particular, @devashishshankar says

I'm not totally sure, but this could be because of the connection pool being maintained by the underlying httplib2 library. httplib2 is holding connection pools, but doesn't seem to be refreshing them

I am fairly certain this is the issue.

[dhermes]

The main problem is likely that httplib2.Http keeps a cache

[devashishshankar]

Yes, seems like that is the issue. Until this gets fixed, is there a temporary workaround I could use? Currently most of my long running jobs fail - and they are also unable to save the state (checkpoint) in datastore because of the same issue. Putting a try catch around every logging statement also doesn't seem feasible.

One hacky solution I can think of is hitting some dummy api call every minute or so in a parallel thread, so the connection remains alive. Perhaps there is some better way?

[dhermes]

See httplib2shim and the discussion in #1214 to just avoid using httplib2 all together

[devashishshankar]

So, I'm passing the gcloud_requests client for logging - and it seems to be working fine for now.

[dhermes]

Glad that helped. I am close to ripping out httplib2 and making it easier to swap out transports. We hope to have requests as the default transport.

@devashishshankar and @znorris Should we leave this issue open or consider it "fixed" (with a bandaid, but still fixed)

[znorris]

@dhermes I haven't had time to test it myself, but from the conversation it sounds fixed to me. I vote for closing. Thanks for the help everyone.

[mauliksoneji]

@dhermes @znorris can you please mention from which version onwards this issue has been fixed?

[tseaver]

@mauliksoneji httplib2 was removed in PR #3674, more than two years ago. google-cloud-datastore 1.3.0 and later use requests instead.

[fucker007]

who can help me.

Traceback (most recent call last):
  File "test_google_translate_translate.py", line 6, in <module>
    word = prepare.translate("I am fucker, we all known something about this word,thers is no fool in the word,but someone get other way to make them believe in it")
  File "/home/nathan/workspace/dmomb/kaggle_dowload_data/google_translate.py", line 16, in translate
    result = self.client.translate(text,target_language='zh')
  File "/home/nathan/.local/lib/python3.6/site-packages/google/cloud/translate_v2/client.py", line 268, in translate
    response = self._connection.api_request(method="POST", path="", data=data)
  File "/home/nathan/.local/lib/python3.6/site-packages/google/cloud/_http.py", line 419, in api_request
    timeout=timeout,
  File "/home/nathan/.local/lib/python3.6/site-packages/google/cloud/_http.py", line 277, in _make_request
    method, url, headers, data, target_object, timeout=timeout
  File "/home/nathan/.local/lib/python3.6/site-packages/google/cloud/_http.py", line 315, in _do_request
    url=url, method=method, headers=headers, data=data, timeout=timeout
  File "/home/nathan/.local/lib/python3.6/site-packages/google/auth/transport/requests.py", line 444, in request
    self.credentials.before_request(auth_request, method, url, request_headers)
  File "/home/nathan/.local/lib/python3.6/site-packages/google/auth/credentials.py", line 133, in before_request
    self.refresh(request)
  File "/home/nathan/.local/lib/python3.6/site-packages/google/oauth2/service_account.py", line 359, in refresh
    access_token, expiry, _ = _client.jwt_grant(request, self._token_uri, assertion)
  File "/home/nathan/.local/lib/python3.6/site-packages/google/oauth2/_client.py", line 153, in jwt_grant
    response_data = _token_endpoint_request(request, token_uri, body)
  File "/home/nathan/.local/lib/python3.6/site-packages/google/oauth2/_client.py", line 105, in _token_endpoint_request
    response = request(method="POST", url=token_uri, headers=headers, body=body)
  File "/home/nathan/.local/lib/python3.6/site-packages/google/auth/transport/requests.py", line 186, in __call__
    six.raise_from(new_exc, caught_exc)
  File "<string>", line 3, in raise_from
google.auth.exceptions.TransportError: ('Connection aborted.', OSError(0, 'Error'))

[tseaver]

Your traceback is unrelated to this issue. Please use the support links to get support.

[FraserTooth]

Had an issue with this occasionally when doing long CREATE OR REPLACE MODEL queries using google.cloud.bigquery.Client()
Appears to be fixed by installing httplib2shim.

I made sure I had the latest version of the google libs, so I dunno why httplib2 is still in there 🤷, judging by the trace it looked like it was in for Python2.7 support.