* Add tests for query callbacks after connection-level errors
* Ensure callbacks are executed for all queued queries after connection-level errors
Separates socket errors from error messages, sends socket errors to all queries in the queue, marks clients as unusable after socket errors.
This is not very pleasant but should maintain backwards compatibility…?
* Always call `handleError` asynchronously
This doesn’t match the original behaviour of the type errors, but it’s correct.
* Fix return value of `Client.prototype.query` in immediate error cases
* Mark clients with closed connections as unusable consistently
* Add tests for error event when connecting Client
* Ensure the promise and callback versions of Client#connect always have the same behaviour
* Give same error to queued queries as to active query when ending
and do so in the native Client as well.
* Restore original ordering between queued query callbacks and 'end' event
Defaults to built-in `tls.checkServerIdentity` method in the event one is not passed into `pgConfig.ssl`
Found breaking in v9.4.2 vs v9.4.1 a la 49054717b4ec0c6d477f04c2becd1f9680b2d13a
cc @tobio @brianc
Existing log code was outputting 'connecting new client' twice and saying 'new client connected', creating a false impression when an error (like a timeout) was present.
This happened to work before because `Query.portalName` was undefined,
but in order to be able to set the portal explicitly it should be using
`Query.portal`.
This commit adds some finer grained detail to handling the postmaster's
response to SSL negotiation packets, by accounting for the possibility
of an 'E' byte being sent back, and emitting an appropriate error.
In the naive case, the postmaster will respond with either 'S' (proceed
with an SSL connection) or 'N' (SSL is not supported). However, the
current if statement doesn't account for an 'E' byte being returned
by the postmaster, where an error is encountered (perhaps unable to
fork due to being out of memory).
By adding this case, we can prevent confusing error messages when SSL is
enforced and the postmaster returns an error after successful SSL
connections.
This also brings the connection handling further in line with
libpq, where 'E' is handled similarly as of this commit:
a49fbaaf8d
Given that there are no longer pre-7.0 databases out in the wild, I
believe this is a safe change to make, and should not break backwards
compatibility (unless matching on error message content).
* Replace if statement with switch, to catch 'S', 'E' and 'N' bytes
returned by the postmaster
* Return an Error for non 'S' or 'N' cases
* Expand and restructure unit tests for SSL negotiation packets
The support for the `noAssert` argument dropped in the upcoming
Node.js v.10.x. All input is then validated no matter if this
is set to true or not. Just remove the argument because of that.
