refactor: improve compilation time for neard tests #4263
Conversation
matklad
requested review from
bowenwang1996,
evgenykuzyakov,
frol,
MaksymZavershynskyi and
SkidanovAlex
as code owners
See near#4027 for discussion why this layout is better. TL;DR is that we don't re-link neard repeatedly. As a cherry on top, we now get to proprely share NodeCluster code between the tests, instead of it being compiled afresh for each test.
Implementation wise, replace a global lazy_static_state with local once_cell state. As a bonus, that gets rid of a Mutex
|
Hehe, turns out that this is not trivial -- there was a bit of a global state, which caused interverence between the tests in the same process. The last commit removes that global state. |
|
@bowenwang1996 or @evgenykuzyakov , could you take another look at this commit: d8d0fd8 ? It's a bit more than "just move the code around", as we need to fix deps between the tests as well. |
| pub fn new(genesis_runtime_config: RuntimeConfig) -> Self { | ||
| Self { | ||
| genesis_runtime_config: Arc::new(genesis_runtime_config), | ||
| with_lower_storage_cost: OnceCell::new(), |
There was a problem hiding this comment.
Alternatively just go ahead and construct the alternative configuration. It would increase memory usage if the protocol feature is not used but simplify things (and minimally speed things up) for the case of the feature being enabled.
There was a problem hiding this comment.
Yeah, that's an extremely reasonable suggestion, to say the least! There's really no need for any kind of laziness here.
For what its worth, LGTM. |
|
@matklad any objections to merging this PR? |
|
This PR has been automatically marked as stale because it has not had recent activity in the 2 weeks. |
|
I still want to merge this, but the specific change no longer is the best way to fix this after https://github.com/near/nearcore/pull/4381/files#r657012158. Let me close this and create an issue instead. |
See #4027 for discussion why this layout is better. TL;DR is that we
don't re-link neard repeatedly.
As a cherry on top, we now get to proprely share NodeCluster code
between the tests, instead of it being compiled afresh for each test.