Rustup (the community package manage for the Rust language) was starting to really suffer : CI times were up at ~ one hour.
We’ve made some strides in bringing this down.
Caching factory for test scenarios
The first thing, which achieved about a 30% reduction in test time was to stop recreating all the test context every time.
Rustup tests the download/installation/upgrade of distributions of Rust. To avoid downloading gigabytes in the test suite, the suite creates mocks of the published Rust artifacts. These mocks are GPG signed and compressed with multiple compression methods, both of which are quite heavyweight operations to perform – and not actually the interesting code under test to execute.
Previously, every test was entirely hermetic, and usually the server state was also unmodified.
There were two cases where the state was modified. One, a small number of tests testing error conditions such as GPG signature failures. And two, quite a number of tests that were testing temporal behaviour: for instance, install nightly at time A, then with a newer server state, perform a rustup update and check a new version is downloaded and installed.
We’re partway through this migration, but compare these two tests:
fn check_updates_some() {
check_update_setup(&|config| {
set_current_dist_date(config, "2015-01-01");
config.expect_ok(&["rustup", "update", "stable"]);
config.expect_ok(&["rustup", "update", "beta"]);
config.expect_ok(&["rustup", "update", "nightly"]);
set_current_dist_date(config, "2015-01-02");
config.expect_stdout_ok(
&["rustup", "check"],
for_host!(
r"stable-{0} - Update available : 1.0.0 (hash-stable-1.0.0) -> 1.1.0 (hash-stable-1.1.0)
beta-{0} - Update available : 1.1.0 (hash-beta-1.1.0) -> 1.2.0 (hash-beta-1.2.0)
nightly-{0} - Update available : 1.2.0 (hash-nightly-1) -> 1.3.0 (hash-nightly-2)
"
),
);
})
}
fn check_updates_some() {
test(&|config| {
config.with_scenario(Scenario::ArchivesV2_2015_01_01, &|config| {
config.expect_ok(&["rustup", "toolchain", "add", "stable", "beta", "nightly"]);
});
config.with_scenario(Scenario::SimpleV2, &|config| {
config.expect_stdout_ok(
&["rustup", "check"],
for_host!(
r"stable-{0} - Update available : 1.0.0 (hash-stable-1.0.0) -> 1.1.0 (hash-stable-1.1.0)
beta-{0} - Update available : 1.1.0 (hash-beta-1.1.0) -> 1.2.0 (hash-beta-1.2.0)
nightly-{0} - Update available : 1.2.0 (hash-nightly-1) -> 1.3.0 (hash-nightly-2)
"
),
);
})
})
}
The former version mutates the date with set_current_dist_date; the new version uses two scenarios, one for the earlier time, and one for the later time. This permits the server state to be constructed only once. On a per-test basis it can move as much as 50% of the time out of the test.
Single binary for the integration test suite
The next major gain was moving from having 14 separate integration test binaries to just one. This reduces the link cost of linking the test binaries, all of which link in the same library. It also permits us to see unused functions in our test support library, which helps with cleaning up cruft rather than having it accumulate.
Hard linking rather than copying ‘rustup-init’
Part of the test suite for each test is setting up an installed rustup environment. Why not start from scratch every time? Well, we obviously have tests that do that, but most tests are focused on steps beyond the new-user case. Setting up an installed rustup environment has a few steps, but particular ones are copying a binary of rustup into the test sandbox, and hard linking it under various names: cargo, rustc, rustup etc.
A debug build of rustup is ~20MB. Running 400 tests means about 8GB of IO; on some platforms most of that IO won’t hit disk, on others it will.
In review now is a PR that changes the initial copy to a hardlink: we hardlink the rustup-init built by cargo into each test, and then hardlink that to the various binaries. That saves 8GB of IO, which isn’t much from some perspectives, but it adds pressure on the page cache, and is wasted work. One wrinkle is a very low max-links limit on NTFS of 1023; to mitigate that we count the links made to rustup-init and generate a new inode for the original to avoid failures happening.
Future work
In GitHub actions this lowers our test time to 19m for Linux, 24m for Windows, which is a lot better but not great.
I plan on experimenting with separate actions for building release artifacts and doing CI tests – at the moment we have the same action do both, but they don’t share artifacts in the cache in any meaningful way, so we can probably gain parallelism there, as well as turning off release builds entirely for CI.
We should finish the cached test context work and use it everywhere.
Also we’re looking at having less integration tests and more narrow close to the code tests.
Code happens 