Use latest watchman #21546
Use latest watchman #21546
Conversation
clauxx
requested review from
alwx,
shivekkhurana,
vkjr,
smohamedjavid and
siddarthkay
|
@vkjr @smohamedjavid can you test if you still get the watchman bug from #16341 on this branch? |
Jenkins BuildsClick to see older builds (18)
|
|
Looks good to me, metro didn't crash for me locally. |
|
component tests do fail though : |
|
@clauxx : component tests pass for me locally after I revert your changes to |
|
I tested this change on Android and iOS and it works well locally.
|
Can confirm that the bug is NOT present anymore, and file changes (hot reload) work well as expected. Thanks 🙏 |
| "ignore_dirs": [ | ||
| "android/build", | ||
| "target" | ||
| "target", | ||
| ".clj-kondo", | ||
| ".shadow-cljs" | ||
| ] |
There was a problem hiding this comment.
while debugging metro, found that watchman is keeping track of these two directories (e.g. .clj-kondo has ~12000 files usually and .shadow-cljs ~3000). On top of that, working in the editor/IDE causes constant changes to .clj-kondo's cache, which re-trigger watchman making it do extra work for no reason (and also hogging up my CPU at times).
The latest metro version requires watchman capabilities that are not available in the version we locked, so it was not used. Instead we used the "node watcher", which is slower and buggier (at least for me metro was timing out quite often)
clauxx
force-pushed
the
cl-fix-metro-issue
branch
from
9b4fe71
to
f380670
Compare
@jakubgs do you mean upgrade watchman from the nix/pkgs file? Could do, but I'm not as versed in nix, so it seemed easier & less time consuming to just use the latest version by removing it tbh. On a sidenote, I removed the yarn.lock/podfile changes as it seems like it wasn't necessary for my issue. It seems like the problem with the un-killable node process was the linux kernel, where the node process was stuck in the "uninterruptible sleep" state, so that should explain why SIGKILL didn't do much. Updating the kernel version fixed the issue so far. cc: @siddarthkay |
There was a problem hiding this comment.
I think we used Nix for the package because - as the README states - we simply wanted to downgrade.
What I don't get is why is the fb-watchman in yarn.lock even older than the 4.9.0 that we provide via Nix right now?
Lines 6053 to 6058 in ca22b67
Summary
fixes #21545
Watchman was locked at an older version due to #16341. After the latest upgrade though, metro requires the latest watchman version and falls back to the "node watcher" (slower and buggier). This ends up (for me) with random timeouts when starting metro, since it takes the node watcher quite a bit of time to go through all the files, and we have a lot of them.
Need to make sure the issue from the latest Watchman, which was mentioned above, doesn't appear for anyone.
Also, I started getting a weird issue with metro's node process not dying after getting closed/killed. Sometimes it lost connection to the device and neither CTRL-C nor SIGKILL could get rid of the rogue
nodeprocess. Inspecting the process' strace, I noticed that it tried to access non-existent files/directories insidenode-modules, right at the moment node became non-responsive. Not sure how that's related to the unkillable process, but re-creating theyarn.lockseems to have fixed it (for now).The strace:
status: ready