reuse connections on table.insert

[alexander-fenster]

TL;DR: the current implementation of BigQuery establishes a new connections on each table.insert() call. This causes problems for Cloud Functions because of connection quota that our users hit. Please update the client libraries to reuse connections.

Why new connections cause problems:

Cloud Functions is a serverless platform for executing snippets of code. Many our customers use function to store events in Big Query. GCF has a quota on connections, and this quota is relatively low because each connection requires a NAT port to be allocated and maintained until TCP packet timeout; and there are only 32k ports per server.

Why BigQuery is worse than other client libraries:

I found three types of clients:

1. Clients that always reuse connections - even if we construct the client object in a local scope - @google-cloud/storage
2. Clients that reuse connectoins only if we declare the client object at global scope - @google-cloud/pubsub, @google-cloud/language, @google-cloud/spanner
3. Clients that never reuse connectoins - @google-cloud/bigquery

It would be nice if all libraries worked as 1.

How to reproduce the problem:

Every call to the function exported by the code below creates a new connection.

******************* function.js ****************

const bigquery = require('@google-cloud/bigquery')();
const table = bigquery.dataset('Tests').table('RandomNumbers');

exports.function = function(req, res) {
  var r = Math.floor(Math.random() * 100);
  console.log("Generated random number " + r);
  var row = {number: r};
  table.insert(row, function(err, apiResponse) {
    if (err) {
      result = 'Error inserting data to bigquery';
      console.log(result + ": " + JSON.stringify(err, null, 2));
      res.status(500).send("Error" + err);
    } else {
      res.status(200).send("" + r);
    }
  });
};

---

************* package.json ****************

{
  "dependencies": {
    "@google-cloud/bigquery": "^0.10.0"
  }
}

---

We also found that BigQuery uses IPv4, while other libraries prefer IPv6. This should also be fixed.

[Googlers: internal tracking id b/68240537]

[stephenplusplus]

I believe what we want is a Keep-Alive header for our requests. This is possible using the npm request module that we use, and is actually the default behavior. However, we turned it off specifically for Google Cloud Functions due to users receiving lots of socket hang up errors (see here and the rest of the discussion in the issue: googleapis/google-cloud-node#2254 (comment)).

[stephenplusplus]

Related: Storage API has poor performance in Google Cloud Functions

[shannb]

Any resolution to this? @alexander-fenster were you able to resolve the issue but setting the forever agent to false on a fork of google-cloud/storage?

I have run into the same issue using bigquery. Is there some part of the google-cloud/bigquery api that also uses google-cloud/storage?

If not @stephenplusplus could you provide a reference for where in the bigquery api repository this update needs to occur?

[stephenplusplus]

If we confirm that this PR was a mistake, we would disable it in the same place, and all APIs would be immediately corrected.

For any users that want to undo the change from that PR, you can use request overrides to turn Keep-Alive back on:

const BigQuery = require('@google-cloud/bigquery')
const bigquery = new BigQuery({...})

bigquery.interceptors.push({
  request: reqOpts => {
    reqOpts.forever = true
    return reqOpts
  }
})

[danoscarmike]

Ping. Next steps?

[emilyofficial]

Any updates? @stephenplusplus

[stephenplusplus]

Sorry, I was hoping @alexander-fenster would advise.

[amineharbaoui]

Any news for this issue ??