Jens Johansen
b99901c433
548bff1was meant to do less. It actually turns out slower though: Before548bff1: [ 07:52 | 100.0% | + 2252 | - 0 ]: weak/variance/unconstrained_inference real 7m58.017s user 11m15.944s sys 2m42.890s => 209.5 ms per test. After548bff1: [ 08:47 | 100.0% | + 2252 | - 0 ]: weak/variance/unconstrained_inference real 8m53.362s user 10m27.086s sys 1m10.925s => 234.0 ms per test. What it did was, that it introduced - I'm guessing this is the reason - some non-monomorphism in the serialization code by having sinks of different types. Running `time out/ReleaseX64/dart --enable-asserts pkg/front_end/test/fasta/weak_suite.dart --traceStepTiming -DupdateExpectations=false -DsemiFuzz=false` (with code changed so it's three shards, running only the first one) I get numbers like this for writing the dill file though: Before548bff1: 0:01:28.369107 After548bff1: 0:01:44.803383 (again this is for 1 out of 3 shards, but illustrates the point --- serialization is slower). This CL changes this to always serialize into the same sink (in memory), then - if needed (as in the original CL) - write it to disk. This gets (again with 1 out of 3 shards) writing the dill down to 0:01:21.872812, and the total run looks like this: [ 07:18 | 100.0% | + 2252 | - 0 ]: weak/variance/unconstrained_inference real 7m23.361s user 10m48.069s sys 1m30.336s => 194.4 ms per test. (rebased run: [ 07:13 | 100.0% | + 2254 | - 0 ]: weak/variance/unconstrained_inference real 7m18.828s user 10m44.163s sys 1m31.777s => 192.1 ms per test.) Change-Id: I2906807bbab407fb499aa910afd20f41347bc3c6 Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/245374 Reviewed-by: Johnni Winther <johnniwinther@google.com> Commit-Queue: Jens Johansen <jensj@google.com>
Package validation
The packages in pkg/ are automatically validated on the LUCI CI bots. The
validation is largely done by the tools/package_deps package; it can be tested
locally via:
dart tools/package_deps/bin/package_deps.dart
Packages which are published
There are several packages developed in pkg/ which are published to pub.
Validation of these packages is particularly important because the pub tools are
not used for these packages during development; we get our dependency versions
from the DEPS file. Its very easy for the dependencies specified in a package's
pubspec file to get out of date wrt the packages and versions actually used.
In order to better ensure we're publishing correct packages, we validate some properties of the pubspec files on our CI system. These validations include:
- that the dependencies listed in the pubspec are used in the package
- that all the packages used by the source are listed in the pubspec
- that we don't use relative path deps to pkg/ or third_party/ packages
Packages which are not published
For packages in pkg/ which we do not intend to be published, we put the following comment in the pubspec.yaml file:
# This package is not intended for consumption on pub.dev. DO NOT publish.
publish_to: none
These pubspecs are still validated by the package validation tool. The contents are more informational as the pubspecs for these packages are not consumed by the pub tool or ecosystem.
We validate:
- that the dependencies listed in the pubspec are used in the package
- that all the packages used by the source are listed in the pubspec
- that a reference to a pkg/ package is done via a relative path dependency