Running `dep ensure` via `make vendor`, with a clean dep cache, takes
around 30 seconds.
For run https://circleci.com/gh/influxdata/platform/1838, we can see
that `make vendor` took 33 seconds and saving the dep cache took 38
seconds. The 38 seconds to save the dep cache is paid the first time
that a particular version of Gopkg.lock is pushed. A subsequent run that
fully uses the dep cache costs 14 seconds to restore the dep cache and 5
seconds to run `make vendor`, so this saves about 10 seconds for any run
that doesn't change Gopkg.lock. See
https://circleci.com/gh/influxdata/platform/1840.
With a clean GOCACHE, `make test-go` took about 30 seconds. It's hard to
tell the exact time since `make vendor` was implicitly executed in the
same block.
Now, `make test-go` uses GOCACHE and takes about 5 seconds to restore
the cache, then about 8 seconds to run through a fully cached set of
tests. If any test is actually broken or doesn't compile, this will give
us fast feedback.
But this change also runs `go test -race -count=1`, so that we fully
exercise any possible data races. That currently takes about 19 seconds.
Then finally we save GOCACHE, which has our `go test` results, and a
populated build cache for our test files, including built with the race
detector. Saving that takes about 14 seconds.
That means we took about 30 seconds originally to just run `go test
-count=1`, and now we take (5+8+19+14)=46 seconds to run plain go test
and go test with the race detector. There's probably an argument to be
made for just running with the race detector, but running both gives us
more coverage IMO, and it does allow us to run tests that aren't enabled
on race builds.
See https://circleci.com/gh/influxdata/platform/1840 for a run with full
caching.
Mark Rushakoff
lukevmorris
Chris Goller
Michael Desa
Jonathan A. Sternberg
Nathaniel Cook
j. Emrys Landivar (docmerlin)