ericl
Follow
Pinned
2,774 contributions in the last year
Less
More
Contribution activity
February 2021
Created 27 commits in 1 repository
Created a pull request in ray-project/ray that received 1 comment
Use more IO workers by default
This is particularly important for restore workers, since there is no batching of requests there. In the future we can consider having a large numb…
+1
−1
•
1
comment
Opened 20 other pull requests in 1 repository
ray-project/ray
18
merged
1
open
1
closed
- [client] Allow ignoring version mismatch with env var for debugging
- Bump the Ray client version
- Auto report object store memory usage; remove some deprecated code
- Handle unhandled exception handler == nullptr in Java
- Fix ray stack
- Fix deadlock in unhandled exception handler and re-merge (#3)
- [WIP] Remove old sched
- Fix OSX error and re-merge unhandled exceptions handling
- [tune] Avoid crash in client mode when return results creating logdir
- Revert "Revert "Unhandled exception handler based on local ref counti…
- Unhandled exception handler based on local ref counting
- Fix up shuffle example script to work in multi-node
- Add tip on how to disable Ray OOM handler
- Use strict equality check for Ray client protocol version
- Check in shuffle code as experimental
- Check Ray client protocol version
- Remove LRU eviction docs
- Fix naming of ray_spilled_objects directory
- Add Ray client protocol version
- Unconditionally retry all RPC errors on client connect
Reviewed 54 pull requests in 1 repository
ray-project/ray 54 pull requests
- Auto report object store memory usage; remove some deprecated code
- [core] Randomize actor ID to avoid collisions
- [Core] Raylet to pick the node manager port
- [core] Spill tasks from waiting queue
- Fixing bug when dispatching tasks to deleted placement group
- Emulate multi node in shuffle.py
- [Core] Fix the issue with duplicated args
- [WIP] File mounts prototype
- Raylet request resource report endpoint
- [Monitor] Log some diagnosis information on startup
- [Object Spilling] Support multiple directories for spilling.
- [Core] Ownership-based Object Directory - Enabled by default
- node_ip changes while sorting the node ips based on last used
- [Core - Autoscaler] Upon autoscaler failure, propagate error message to all current and future drivers
- [Core] Make Actor DEAD and Save Exceptions in GCS When Error Happens in Constructor
- [Test] Add multi node data processing cluster.yaml
- [core] Fix bugs in admission control
- Revert "Fix OSX error and re-merge unhandled exceptions handling (#14…
- [Shuffle] Use progress bar for experimental.shuffle
- [ray_client]: Skip flaky test_cancel_chain on Windows
- [Object Spilling] Fix the race condition.
- Revert "[Ownership based object directory] Turn on by default. (#13964)"
- Revert "[Core]Fix ray.kill doesn't cancel pending actor bug"
- [Ownership based object directory] Turn on by default
- Revert "Revert "Revert "Unhandled exception handler based on local ref counti…"
- Some pull request reviews not shown.
Created an issue in ray-project/ray that received 12 comments
Too many workers launched in multi node shuffle benchmark.
On a 5 node machine with 1GB object store memory each:
python -m ray.experimental.shuffle --num-partitions=500 --partition-size=20e6 --ray-address=…
12
comments
Opened 9 other issues in 1 repository
ray-project/ray
8
open
1
closed
- Require opt-in to switching plasma to /tmp instead of /dev/shm
- ray memory --stats-only should not fetch all reference table info
- Autoscaler stopping nodes when updating cluster
- Excessive heap memory usage in raylet / owner process when shuffling many objects
- Autoscaling should not remove nodes with active or spilled objects
- [RFC] loading packaged ray functions
- [RFC] k8s-native worker pool
- Move private APIs to ray._private
- Certain types of errors aren't deserialized properly in Ray client
49
contributions
in private repositories
Feb 1 – Feb 26