commit 3d6432c72ee9bfb55444a618b3b161bb1c0c6c7a
author Jonathan Amsterdam <jba@google.com> Wed Jun 02 11:35:39 2021 -0400
committer Jonathan Amsterdam <jba@google.com> Fri Jun 04 13:12:34 2021 +0000
tree 900d9049d3189e636432e94050bcc54d09bec095
parent 50c1ec4d2c83944e0f6db9611f844a4ab66591e6

event: add a namespace to every event

Add a namespace to every event.

It can be set with Builder.Namespace.

By default we use the import path of the package that builds the
event.

The Exporter memoizes the map from pcs to namespaces. This saves both
the time to extract the frame and parse the fully-qualified function
name, and the allocation that is inevitable because
runtime.CallersFrames takes a slice.

This is significantly slower in the Noop cases (see below), so the
user must enable it with an option.

name                  old time/op    new time/op    delta
LogEventNoExporter-8     616ns ± 2%     612ns ± 2%    -0.67%  (p=0.019 n=17+20)
LogEventNoop-8          1.83µs ± 3%   13.51µs ± 4%  +638.00%  (p=0.000 n=18+19)
LogEventDiscard-8       10.7µs ± 2%    23.7µs ± 2%  +121.35%  (p=0.000 n=20+17)
LogEventfDiscard-8      16.6µs ± 2%    28.8µs ± 1%   +73.47%  (p=0.000 n=19+18)
TraceEventNoop-8        41.5µs ± 3%    78.6µs ± 1%   +89.63%  (p=0.000 n=19+18)
MetricEventNoop-8       16.4µs ± 2%    40.9µs ± 2%  +149.97%  (p=0.000 n=17+20)
Logrus-8                55.4µs ± 2%    54.3µs ± 3%    -1.97%  (p=0.000 n=19+17)
Logrusf-8               46.4µs ± 3%    45.1µs ± 1%    -2.82%  (p=0.000 n=20+18)
Baseline-8               280ns ± 2%     278ns ± 2%    -0.72%  (p=0.025 n=19+20)
LogStdlib-8             8.54µs ± 1%    8.49µs ± 2%    -0.60%  (p=0.043 n=18+18)
LogPrintf-8             13.1µs ± 2%    13.0µs ± 2%    -0.72%  (p=0.010 n=18+18)
LogZap-8                20.4µs ± 3%    20.1µs ± 1%    -1.43%  (p=0.000 n=20+18)
LogZapf-8               21.7µs ± 3%    21.6µs ± 1%      ~     (p=0.797 n=20+19)
LogZerolog-8            7.47µs ± 2%    7.40µs ± 1%    -0.87%  (p=0.006 n=18+19)
LogZerologf-8           12.5µs ± 3%    12.6µs ± 2%      ~     (p=0.171 n=18+19)

Further testing showed that the time is all due to runtime.Callers;
caching the pcs is very effective at eliminating any overhead from
extracting frame information.  (Note that TraceEventNoop and
MetricEventNoop are misnamed; they install a printer that writes to
io.Discard.)

Change-Id: I37bb5e6194be6dcaaa471086dcd3d4850bfd531c
Reviewed-on: https://go-review.googlesource.com/c/exp/+/324291
Trust: Jonathan Amsterdam <jba@google.com>
Run-TryBot: Jonathan Amsterdam <jba@google.com>
Reviewed-by: Ian Cottrell <iancottrell@google.com>