Experiment, Trial, and config parsing moves into an `experiment` package.
Notably, the new public facing APIs will be
```
from ray.tune.experiment import Experiment
from ray.tune.experiment import Trial
```
Users often have UDFs that require arguments only known at Dataset creation time, and sometimes these arguments may be large objects better suited for shared zero-copy access via the object store. One example is batch inference on a large on-CPU model using the actor pool strategy: without putting the large model into the object store and sharing across actors on the same node, each actor worker will need its own copy of the model.
Unfortunately, we can't rely on closing over these object refs, since the object ref will then be serialized in the exported function/class definition, causing the object to be indefinitely pinned and therefore leaked. It's much cleaner to instead link these object refs in as actor creation and task arguments. This PR adds support for threading such object refs through as actor creation and task arguments and supplying the concrete values to the UDFs.
In [PR](https://github.com/ray-project/ray/pull/24764) we move the reconnection to GcsRPCClient. In case of a GCS failure, we'll queue the requests and resent them once GCS is back.
This actually breaks request with timeout because now, the request will be queued and never got a response. This PR fixed it.
For all requests, it'll be stored by the time it's supposed to be timeout. When GCS is down, we'll check the queued requests and make sure if it's timeout, we'll reply immediately with a Timeout error message.
Ray (on K8s) fails silently when running out of disk space.
Today, when running a script that has a large amount of object spilling, if the disk runs out of space then Kubernetes will silently terminate the node. Autoscaling will kick in and replace the dead node. There is no indication that there was a failure due to disk space.
Instead, we should fail tasks with a good error message when the disk is full.
We monitor the disk usage, when node disk usage grows over the predefined capacity (like 90%), we fail new task/actor/object put that allocates new objects.
Task/actor/object summary
Tasks: Group by the func name. In the future, we will also allow to group by task_group.
Actors: Group by actor class name. In the future, we will also allow to group by actor_group.
Object: Group by callsite. In the future, we will allow to group by reference type or task state.
Enable checking of the ray core module, excluding serve, workflows, and tune, in ./ci/lint/check_api_annotations.py. This required moving many files to ray._private and associated fixes.
Uses the async KV API for downloading in the runtime env agent. This avoids the complexity of running the runtime env creation functions in a separate thread.
Some functions are still sync, including the working_dir/py_modules upload, installing wheels, and possibly others.
This PR adds a format-based file extension path filter for file-based datasources, and sets it as the default path filter. This will allow users to point the read_{format}() API at directories containing a mixture of files, and ensure that only files of the appropriate type are read. This default filter can still be disabled via ray.data.read_csv(..., partition_filter=None).
This allows correct logging of tuple entries in configs, e.g. PolicySpec (which is a namedtuple) from multiagent.policies key. Without this, the whole PolicySpec is serialized as a string, which doesn't allow to filter run by specific key from this tuple.
I’d like to propose a bit changes to the API. Currently we are returning the dict of ID -> value mapping when the list API is returned. But I am thinking to change this to a list because the sort will become ineffective if we return the dictionary. So, it’s ideal we use the list to keep the order (it’s important for deterministic order)
Also, for some APIs, each entry doesn’t have a unique id. For example, list objects will have duplicated object IDs from their entries, which is not working with dict return type (e.g., there can be more than 1 Object ID entry if the object is locally referenced & borrowed by task/pinned in memory)
Also, users can easily build dict index on their own if it is necessary.
The error message in #25638 indicates we should use protobuf>3.19.0 to generated code so that we can work with python protobuf >= 4.21.1. Try generating wheels to see if this works.
In this PR, we integrate GCS fault tolerance with ray serve.
- Add timeout with 5s for kv.
Rollback should be added to all methods, which will come after.
Basic testing for KV timeout in serve and deploy is added.
Reduce stage in push-based shuffle fails to complete at 100k output partitions or more. This is likely because of driver or raylet load from having too many tasks in flight at once.
We can fix this from ray core too, but for now, this PR adds pipelining for the reduce stage, to limit the total number of reduce tasks in flight at the same time. This is currently set to 2 * available parallelism in the cluster. We have to pick which reduce tasks to submit carefully since these are pinned to specific nodes. The PR does this by assigning tasks round-robin according to the corresponding merge task (which get spread throughout the cluster).
In addition, this PR refactors the map, merge, and reduce stages to use a common pipelined iterator pattern, since they all have a similar pattern of submitting a round of tasks at a time, then waiting for a previous round to finish before submitting more.
Related issue number
Closes#25412.
Consolidate tune/train report/checkpoint functionality by working with a unified Session interface.
The goal of this PR is to establish a solid Session and Session.report path.
In favor of having less merging conflict (as other folks are doing the whole package renaming) and control the scope of this PR, I have intentionally left out some migration. More PRs to follow. Feel free to comment on the ideal final state.
To give an idea of the final directory structure. This is a for 2-worker DP training.
```
├── TensorflowTrainer_ce44d_00000_0_2022-06-15_14-40-42
│ ├── checkpoint_000000
│ │ ├── _current_checkpoint_id.meta.pkl
│ │ ├── _preprocessor.meta.pkl
│ │ ├── _timestamp.meta.pkl
│ │ ├── assets
│ │ ├── keras_metadata.pb
│ │ ├── saved_model.pb
│ │ └── variables
│ │ ├── variables.data-00000-of-00001
│ │ └── variables.index
│ ├── events.out.tfevents.1655329242.xw
│ ├── params.json
│ ├── params.pkl
│ ├── progress.csv
│ ├── rank_0
│ │ └── my_model
│ │ ├── assets
│ │ ├── keras_metadata.pb
│ │ ├── saved_model.pb
│ │ └── variables
│ │ ├── variables.data-00000-of-00001
│ │ └── variables.index
│ ├── rank_1
│ │ └── my_model
│ │ ├── assets
│ │ ├── keras_metadata.pb
│ │ ├── saved_model.pb
│ │ └── variables
│ │ ├── variables.data-00000-of-00001
│ │ └── variables.index
│ └── result.json
├── basic-variant-state-2022-06-15_14-40-42.json
├── experiment_state-2022-06-15_14-40-42.json
├── trainable.pkl
└── tuner.pkl
```
Update:
1. Updated a few classes to be backward compatible - while legacy ray train deprecation is ongoing.
2. Marked all places in 1 using "# TODO(xwjiang): Legacy Ray Train trainer clean up!". So we can easily clean those up once Antoni's work is landed.
3. All CI and release tests are passing.
Co-authored-by: Eric Liang <ekhliang@gmail.com>
Simplify isort filters and move it into isort cfg file.
With this change, isort will not longer apply to diffs other than to files that are in whitelisted directory (isort only supports blacklist so we implement that instead) This is much simpler than building our own whitelist logic since our formatter runs multiple codepaths depending on whether it is formatting a single file / PR / entire repo in CI.
This uses the generators introduced in #25247 to reduce memory usage during the merge stage in push-based shuffle. These tasks merge groups of map outputs, so it fits a generator pattern where we want to return merged outputs one at a time. Verified that this allows for merging more/larger objects at a time than the current list-based version.
I also tried this for the map stage in random_shuffle, but it didn't seem to make a difference in memory usage for Arrow blocks. I think this is probably because Arrow is already doing some zero-copy optimizations when selecting rows?
Also adds a new line to Dataset stats for memory usage. Unfortunately it's hard to get an accurate reading of physical memory usage in Python and this value will probably be an overestimate in a lot of cases. I didn't see a difference before and after this PR for the merge stage, for example. Arguably this field should be opt-in. For 100MB partitions, for example:
```
Substage 0 read->random_shuffle_map: 10/10 blocks executed
* Remote wall time: 1.44s min, 3.32s max, 2.57s mean, 25.74s total
* Remote cpu time: 1.42s min, 2.53s max, 2.03s mean, 20.25s total
* Worker memory usage (MB): 462 min, 864 max, 552 mean
* Output num rows: 12500000 min, 12500000 max, 12500000 mean, 125000000 total
* Output size bytes: 101562500 min, 101562500 max, 101562500 mean, 1015625000 total
* Tasks per node: 10 min, 10 max, 10 mean; 1 nodes used
Substage 1 random_shuffle_reduce: 10/10 blocks executed
* Remote wall time: 1.47s min, 2.94s max, 2.17s mean, 21.69s total
* Remote cpu time: 1.45s min, 1.88s max, 1.71s mean, 17.09s total
* Worker memory usage (MB): 462 min, 1047 max, 831 mean
* Output num rows: 12500000 min, 12500000 max, 12500000 mean, 125000000 total
* Output size bytes: 101562500 min, 101562500 max, 101562500 mean, 1015625000 total
* Tasks per node: 10 min, 10 max, 10 mean; 1 nodes used
```
## Checks
- [ ] I've run `scripts/format.sh` to lint the changes in this PR.
- [ ] I've included any doc changes needed for https://docs.ray.io/en/master/.
- [ ] I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
- Testing Strategy
- [ ] Unit tests
- [ ] Release tests
- [ ] This PR is not tested :(
Co-authored-by: Eric Liang <ekhliang@gmail.com>
Followup from #24622. This is another step towards pluggability for runtime_env. Previously some plugin classes had `get_uri` which returned a single URI, while others had `get_uris` which returned a list. This PR makes all plugins use `get_uris`, which simplifies the code overall.
Most of the lines in the diff just come from the new `format.sh` which sorts the imports.
Fixes a bug in push-based shuffle in computing the merge task <> reduce tasks mapping when the number of reduce tasks does not divide evenly by the number of merge tasks. Previously, if there were N reduce tasks for one merge task, we would do:
[N + 1, N + 1, ..., N + 1, all left over tasks]
which could lead to negative many reduce tasks n the last merge partition.
This PR changes it to:
[N + 1, N + 1, ..., N + 1, N, N, N, ...]
Related issue number
Closes#25863.
This PR ensures that GCS keeps the IP and PID information about a job so that it can be used to find the job's logs in the dashboard after the job terminates.
@alanwguo will handle any dashboard work in a separate PR.
Co-authored-by: Alex Wu <alex@anyscale.com>
pyarrow.json.ReadOptions are not picklable until Arrow 8.0.0, which we do not yet support. This PR adds a custom serializer for this type and ensures that said serializer is registered before each Ray task submission.
Previously it was not possible to set the object spill config from the RAY_xxx environment variable, unlike other system configs. This is because the config is initialized in Python first, before the C++ config is parsed.
