Adds a set_max_concurrency method to the Searcher API. This method allows for the ConcurrencyLimiter to override the max_concurrency value on searchers with custom internal logic for limiting concurrency (atm. SigOpt and HEBO). This PR also changes the initialisation of SigOpt and HEBO optimisers to happen on set_search_properties instead of in the constructor, so that the new max_concurrency is respected.
Furthermore, this PR breaks up test_tune_restore.py into test_tune_restore_warm_start.py and test_tune_restore.py to deal with a timeout, and ensures that the automatic application of ConcurrencyLimiter in tune.run doesn't override a user-defined ConcurrencyLimiter.
Adds a working failure test for streaming and non-streaming shuffle, without lineage reconstruction. This does a few things.
Test improvements:
- modifies AutoscalingCluster to allow passing an idle node timeout (the default is very low)
- some small improvements to the NodeKiller actor to hopefully improve flakiness.
Shuffle fixes:
- modifies shuffle tracker to wait on futures instead of having tasks signal. During failures, tasks may never signal the tracker, so we can't rely on these to track progress.
Core fixes:
- raylet will exit immediately if it receives the Shutdown RPC with graceful=False - there was a bug here where it's supposed to exit after replying to the client, but the gRPC server goes down for an unknown reason and the client reply is never sent
- On reference deletion, the owner now publishes an additional message to subscribers that the object has been deleted. Previously, this was causing a hang in streaming shuffle because the raylets pulling an object subscribed after the object was already deleted, so they never received the error signal.
In `sample_boundaries`, naive concatenation with `np.concatenate()` doesn't work when the single-column sample blocks have varying lengths (e.g., when the original dataset had non-uniform blocks). This PR fixes this by delegating concatenation and NumPy array conversion to the block builder and block accessor, respectively.
Using Ray pubsub for publishing and subscribing logs via GCS, from Python worker, log importer, dashboard and unit tests.
This change is guarded behind the RAY_gcs_grpc_based_pubsub feature flag.
This PR adds support for publishing and subscribing to logs in Python via GCS pubsub. It also refactors the Python threaded subscriber to support subscribing and calling `close()` from multiple threads.
We can also move tests and logging support to another PR, but it will make the purpose of the refactoring seems less obvious.
Arbitrary API access is pretty rampant at the moment. It is pretty hard to correct it in one go. This is a necessary incremental step towards a cleaner API.
Moving debug_state.txt to the log directory. This will help us finding debug_state.txt from the dashboard. See below.
Add debug_state_gcs.txt. This will display GCS' debug state. GCS will also dump debug state to the file every 10 seconds
For periodic printing of debug state, I made it happen every 1 minute. This is because every 10 seconds usually is very spammy.
This reverts commit f13c2a5350.
Re-land remove PG caching logic.
As a result, pbt scheduler cannot stop and start trial within itself for weight transfer and perturbation now. So these are some changes to pbt scheduler:
1. the trial being perturbed is always left in a PAUSED state upon exiting on_trial_result. This is because instead of maintaining two separate paths for replacing a trial, we consolidate to always "stop" and "restore" and rely on reuse_actor as an optimization if available. (see 2)
2. consolidates pbt replacing a trial with reuse_actor.
3. introduces a NOOP scheduler decision to indicate that (pbt) scheduler has finished its interaction with executor and thus no decision is further needed in Tune loop.
Long term, we should control the interface between scheduler and executor. For example, on_trial_result taking in the whole runner is too much API exposure that we want to remove.
This PR includes the precise reason why actor is dead to `ActorTable`. The `death cause` stored in the table will be propagated to core worker through pubsub, so that core worker can eventually raise a good error message with metadata.
This PR is mostly for implementing "fixture" for nightly test. Note that the current fixture implementation is not that great, and we can probably improve this in the future after refactoring e2e.py.
This fixes slow lazy block evaluation by adding an explicit get_blocks() bulk method, and using that when-ever lazy iteration is not needed.
The root cause of the slowdown was because block splitting requires ray.get() during iteration over block refs, to materialize split blocks. However, this interferes with exponential rampup.
Instead of wrapping the whole training run in a remote call, we only query the files on the node in a remote call. XGBoost-Ray is then started from the local node.
block splitting and makes it off by default. This makes it easier to debug problems potentially related to this feature. Criteria for enabling by default:
- We're confident all nightly tests pass (currently, there may be an issue with large-scale groupby with block splitting).
- We're confident lineage-based reconstruction can work with block splitting.
This test seems to be flaking since ray stop sometimes fails when sending SIGTERM only. While that's desirable to fix, the test is still testing the intended behavior even if we send SIGKILL.
This PR introduces a TrialCheckpoint class which is returned e.g. by ExperimentAnalysis.best_checkpoint. The class enables easy access to cloud storage locations (rather than just local directories before). It also comes with utilities to download, upload, and save trial checkpoints to local and cloud targets.
Running ray status with the changes from #20359
while running an autoscaler older than those changes
results in an error on input "head_ip" to LoadMetricsSummary.
See #20359 (comment)
This PR fixes the bug by restoring head_ip as an optional parameter of LoadMetricsSummary.
non_terminated_nodes calls are expensive for some node provider implementations.
This PR refactors autoscaler._update() such that it results in at most one non_terminated_nodes call.
Conceptually, the change is that the autoscaler only needs a consistent view of the world once per update interval.
The structure of an autoscaler update is now
call non_terminated_nodes to update internal state
update autoscaler status strings
terminate nodes we don't need, removing them from internal state as we go
run node updaters if needed
get nodes to launch based on internal state
There's a small operational difference introduced:
Previously -- After a node is created, its NodeUpdater thread is initiated immediately.
Now -- After a node is created, its NodeUpdater thread is initiated in the next autoscaler update.
This typically will not introduce latency, since the time to get SSH access (a few minutes) is much longer than the autoscaler update interval (5 seconds by default).
Along the way, I've removed the local_ip initialization parameter for LoadMetrics because it was confusing and not useful (and caused some tests to fail)
## Why are these changes needed?
In python, redis rpush is used to broadcast and store the keys. In this PR, we use gcs kv to store the keys. pubsub is still using redis which need to be remove later.
The protocol before this PR:
- worker subscribe to redis key spaces
- worker write the key of function/actor to (export:sqn, key)
- so the other worker will be notified and start to load the data by checking export:sqn
This depends on redis for both kv and pubsub, and this PR fix the kv part.
After this PR:
- worker subscribe to redis key space
- For exporting:
- worker will find the first key not being hold. This is guaranteed by internal kv which right now is a single thread, atomic db. The worker will just check until it find one key not existing and write it (this is single operation). One optimization right now is to use the import counter as the start offset since this counter means all keys before the counter has already been used.
- worker will then write a dummy key to redis key space for broadcasting
- For importer
- It's working as before, but instead of reading from redis, it will read from gcs kv.
This is part in redis removal project.
## Related issue number
https://github.com/ray-project/ray/issues/19443
Remerging #19789 with some fixes for Dask-on-Ray 1TB sort:
- Fixes a bug where the timer was not getting reset correctly
- Increased timeout to 10min just to be safe
- Changed the error to a unique exception ObjectFetchTimedOutError to improve debugging.
This exception should usually indicate a system-level bug.
