Turn on incremental memo by default

[rgrinberg]

For some reason this breaks error reporting with RPC (it now reports stale errors). Investigating as to what the cause is.

[rgrinberg]

Actually, CI demonstrates that the issue is present only on MacOS. I narrowed it down to the updated source not being copied to _build after running the build again. @aalekseyev anyway to make the test to wait for fsevents to pick up the outdated file before starting the build again?

[aalekseyev]

@rgrinberg, "waited for inotify sync" is supposed to correspond to exactly that: the idea is to wait until all file notification events up until that point are processed.
This works by modifying a special file on the filesystem (see special_file_for_inotify_sync in dune_file_watcher.ml) and waiting for that file modification to propagate through file notification system. This relies on events being reported in-order, which seems to be the case for inotify. Maybe that doesn't hold for fswatch on a Mac?

My first thought was that maybe the Sync event doesn't happen at all (e.g. we don't watch the special file), but if that was the case then I think you'd just see a deadlock, instead.

[ghost]

It seems that fsevents supports a "flush" operation that does the same as the inotify sync mechanism: https://developer.apple.com/documentation/coreservices/1445629-fseventstreamflushsync?language=objc

[rgrinberg]

Yeah, that looks useful but it would require us to abandon fswatch on MacOS and switch to fsvents directly.

…

On Aug 31, 2021, 7:00 AM -0700, Jérémie Dimino ***@***.***>, wrote: It seems that fsevents supports a "flush" operation that does the same as the inotify sync mechanism: https://developer.apple.com/documentation/coreservices/1445629-fseventstreamflushsync?language=objc — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[ghost]

Indeed. In the long run that seems inevitable if we want proper fine grained file watching. fswatch/inotifywait were an easy way to bootstrap the system but they don't give us much control. For instance, we can't dynamically add watches with fswatch, so we can't watch external files and directories while we can by using the inotify or fsevents API directly.

[ghost]

@rgrinberg can you disable the offending tests on OSX for now and merge this PR? And while you are at it, delete the environment variable and the old mode entirely.

The new mode is usable enough now and there is going to be a fair amount of polishing to do before it's ready for release. So we might as well make the new mode the new reality ASAP.

[rgrinberg]

Is there a way to do enabled_if on inline tests? I'm not sure how to disable tests on OSX.

[ghost]

Putting it on the whole library stanza should work. I'm trying that.