Ports used by the unit tests have been allocated by pulling them out of a list, with no guarantee that the port is not actually in use. This central allocation point also means that tests cannot be split into separate packages to be run in parallel, as the ports allocated between the various packages will be allocated multiple times and end up intermittently clashing.
There is also no guarantee, even when the tests are run serially, that the ports will not clash with services already running on the machine.
This patch (largely) replaces the use of this centralised port allocation with pre-created listeners injected into the test via the file descriptor import mechanism use by Teleport to pass open ports to child processes.
There are still some cases where the old port allocation system is still in use. I felt this was already getting beyond the bounds of sensibly reviewable, so I have left those for a further PR after this.
See-Also: #12421
See-Also: #14408
Linting for non-go files was accidentally dropped in the transition to
GCB (sorry!). This patch restores linting for non-go files and fixes
any lint failures that have crept in during the interim.
Actually tracking down the cause of a failure in the integration tests can
be hard:
* It's hard to get an overall summary of what failed
* The tests sometimes emit no output before timing out, meaning any
diagnostic info is lost
* The emitted logs are too voluminous for a human to parse
* The emitted logs can present information out of order
* It's often hard to tell where the output from one test ends
and the next one begins
This patch attempts to address these concerns without attempting to rewrite
any of the underlying teleport logging.
* It improves the render-tests script to (optionally) report progress per-
test, rather than on a per-package basis. My working hypothesis on the
tests that time out with no output is that go test ./integration is
waiting for the entire set of integration tests tests to be complete
before reporting success or failure. Reporting on a per-test cycle gives
faster feedback and means that any timed-out builds should give at least
some idea of where they are stuck.
* Adds the render-tests filter to the integration and integration-root make
targets. This will show an overall summary of test results, as well as
- Discarding log output from passing tests to increase signal-to-noise
ratio, and
- Strongly delimiting the output from each failed test, making failures
easier to find.
* Removes the notion of a failure-only logger in favour of post-processing
the log events with render-tests. The failure-only logger catches log
output from the tests and only forwards it to the console if the test
fails. Unfortunately, not all log output is guaranteed to pass through
this logger (some teleport packages do not honour the configured logger,
and reports from the go race detector certainly don't), meaning some
output is presented at the time it happens, and other output is batched
and displayed at the end of the test. This makes working out what
happened where harder than it need be.
In addition, this patch also promotes the render-tests script into a fully-
fledged program, with appropriate makefile targets, make clean support, etc.
It is now also more robust in the face on non-JSON output from go test
(which happens if a package fails to compile).
Trent Clarke