ray/doc/source/ray-air/check-ingest.rst

.. _air-ingest:

Configuring Training Datasets
=============================

This page is a guide for setting up, understanding, and diagnosing data ingest problems in Ray AIR.

Data ingest is the process of loading data from storage, apply preprocessing steps, and feeding the data into Trainers in AIR.
For datasets that are small, ingest is usually not an issue. However, ingest can be tricky to set up when datasets grow
large enough so that they may not fit fully in memory on a single node, or even in aggregate cluster memory.

AIR builds its ingest pipeline on :ref:`Ray Datasets <datasets>`, which is a framework-agnostic distributed data loading library. If you have
an existing ingest pipeline (e.g., based on TF data), there is some upfront effort porting your loading code to Datasets.
In return, AIR provides portability across ML frameworks as well as advanced capabilities such as global random shuffles,
which are not possible in less general ML data preprocessing libraries.

Ingest Basics
-------------

.. _ingest_basics:

The following figure illustrates a simple Ray AIR training job that (1) loads parquet data from S3, (2) applies a simple
user-defined function to preprocess batches of data, and (3) runs an AIR Trainer with the given dataset and preprocessor.

.. figure:: images/ingest.svg

..
  https://docs.google.com/drawings/d/1FPta-TvaL7TQoSRp5ofts0dru6mw1Xes-35m5XuqCBc/edit

Let's walk through the stages of what happens when ``Trainer.fit()`` is called.

**Read**: First, AIR will read the Dataset into the Ray object store by calling ``ds.fully_executed()`` on the datasets
that you pass to the Trainer. Dataset blocks that don't fit into memory will be spilled to disk. Note that when you create
the dataset initially, typically only the first block and block metadata is read into memory. The rest of the blocks are
not loaded until ``fit`` is called.

**Preprocessing**: Next, if a preprocessor is defined, AIR will by default ``fit`` the preprocessor (e.g., compute statistics) on the
``"train"`` dataset, and then ``transform`` all given datasets with the fitted preprocessor. This is done by calling ``prep.fit_transform()``
on the train dataset passed to the Trainer, followed by ``prep.transform()`` on remaining datasets. Preprocessors use Dataset APIs to execute
preprocessing in a parallelized way across the cluster. Both read and preprocessing stages use Ray tasks under the hood.

**Training**: Finally, AIR passes a reference to the preprocessed dataset to Train workers (Ray actors) launched by the Trainer. Each worker then
typically calls ``iter_batches``, ``to_tf``, or ``to_torch`` to iterate over the dataset reader retrieved by ``get_dataset_shard``.
These read methods load blocks of the dataset into the local worker's memory in a streaming fashion, only fetching / prefetching a
limited number of blocks at once. Workers loop over the dataset blocks repeatedly until training completes.

Configuring Ingest Per-Dataset
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

It is common to customize processing per-dataset. For example, you may want to enable sharding
on a validation dataset, disable preprocessing of an auxiliary dataset, or adjust ingest strategy per dataset.

Each DataParallelTrainer has a default per-dataset config given by a ``Trainer._dataset_config`` class field. It is a mapping
from dataset names to ``DatasetConfig`` objects, and implements the default behavior described in :ref:`Ingest Basics <ingest_basics>`:

.. code:: python

    # The default DataParallelTrainer dataset config, which is inherited
    # by sub-classes such as TorchTrainer, HorovodTrainer, etc.
    _dataset_config = {
        # Fit preprocessors on the train dataset only. Split the dataset
        # across workers if scaling_config["num_workers"] > 1.
        "train": DatasetConfig(fit=True, split=True),
        # For all other datasets, use the defaults (don't fit, don't split).
        # The datasets will be transformed by the fitted preprocessor.
        "*": DatasetConfig(),
    }

These configs can be overriden via the ``dataset_config`` kwarg, which is recursively merged with the Trainer defaults.
Here are some examples of configuring Dataset ingest options and what they do:

.. tabbed:: Split All

    This example shows overriding the split config for the "valid" and "test" datasets. This means that
    both the valid and test datasets here will be ``.split()`` across the training workers.

    .. literalinclude:: doc_code/air_ingest.py
        :language: python
        :start-after: __config_1__
        :end-before: __config_1_end__

.. tabbed:: Disable Transform

    This example shows overriding the transform config for the "side" dataset. This means that
    the original dataset will be returned by ``.get_dataset_shard("side")``.

    .. literalinclude:: doc_code/air_ingest.py
        :language: python
        :start-after: __config_2__
        :end-before: __config_2_end__


Bulk vs Streaming Reads
-----------------------

Bulk Ingest
~~~~~~~~~~~

By default, AIR loads all Dataset blocks into the object store at the start of training. This provides the best performance if the
cluster has enough aggregate memory to fit all the data blocks in object store memory, or if your preprocessing step is expensive
and you don't want it to be re-run on each epoch. Note that data often requires more space
when loaded uncompressed in memory than when resident in storage.

If there is insufficient object store memory, blocks may be spilled to disk during reads or preprocessing. Ray will print log messages
if spilling is occuring, and you can check this as well with the ``ray memory --stats-only`` utility. If spilling is happening, take
care to ensure the cluster has enough disk space to handle the spilled blocks. Alternatively, consider using machine with more memory /
more machines to avoid spilling.

In short, use bulk ingest when:
 * you have enough memory to fit data blocks in cluster object store;
 * your preprocessing step is expensive per each epoch; and
 * you want best performance when both or either the above conditions are met.

Streaming Ingest (experimental)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

AIR also supports streaming ingest via the DatasetPipeline feature. Streaming ingest is preferable when you are using large datasets
that don't fit into memory, and prefer to read *windows* of data from storage to minimize the active memory required for data ingest.
Note that streaming ingest will re-execute preprocessing on each pass over the data. If preprocessing is a bottleneck, consider
using bulk ingest instead for better performance.

To enable streaming ingest, set ``use_stream_api=True`` in the dataset config. By default, this will configure streaming ingest with a window
size of 1GiB, which means AIR will load ~1 GiB of data at a time from the datasource.
Performance can be increased with larger window sizes, which can be adjusted using the ``stream_window_size`` config.
A reasonable stream window size is something like 20% of available object store memory. Note that the data may be larger
once deserialized in memory, or if individual files are larger than the window size.

If the window size is set to -1, then an infinite window size will be used. This case is equivalent to using bulk loading
(including the performance advantages of caching preprocessed blocks), but still exposing a DatasetPipeline reader.

In short, use streaming ingest when:
 * you have large datasets that don't fit into memory;
 * you want to process small chunks or blocks per window;
 * you can use small windows with small data blocks minimizing or avoiding memory starvation or OOM errors; and
 * your preprocessing step is not a bottleneck or not an expensive operation since it's re-executed on each pass over the data.

.. warning::

    Streaming ingest only applies to preprocessor transform, not preprocessor fitting.
    This means that the preprocessor will be initially fit in bulk, after which data will be transformed
    as it is loaded in a streaming manner.

Reading Data
~~~~~~~~~~~~

The ``get_dataset_shard`` method returns a reader object that is either a ``Dataset`` or ``DatasetPipeline``, depending on whether the ``use_stream_api``
option is set. The former is a finite set of records, and the latter represents an infinite stream of records.
See the following examples for clarification:

.. tabbed:: Bulk Ingest

    This example shows bulk ingest (the default). Data is bulk loaded and made available
    directly via a ``Dataset`` object that can be looped over manually.

    .. literalinclude:: doc_code/air_ingest.py
        :language: python
        :start-after: __config_4__
        :end-before: __config_4_end__

.. tabbed:: Streaming Ingest

    This example shows enabling streaming ingest for the "train" dataset with a *N-byte* window.
    This means that AIR will only load *N* bytes of data from the datasource at a time (the data
    may be larger once deserialized in memory or if individual files are larger than the window).

    .. literalinclude:: doc_code/air_ingest.py
        :language: python
        :start-after: __config_5__
        :end-before: __config_5_end__

Shuffling Data
~~~~~~~~~~~~~~

AIR offers several options for per-epoch shuffling, including *local
(per-shard) shuffling*  and *global (whole-dataset) shuffling*.

.. tabbed:: Local Shuffling

    Local shuffling is an in-iterator shuffle that fills a trainer-local in-memory shuffle
    buffer with records and then pops random samples as batches, keeping the buffer above a
    user-provided threshold to ensure samples are mixed throughout the entirety of the
    trainer's shard. This local shuffle doesn't mix samples across trainer shards between
    epochs as the global shuffle does, and will therefore be a lower-quality shuffle;
    however, since this shuffle only involves a local in-memory buffer, it is much less
    expensive.

    For configuring the size of the in-memory shuffle buffer, it is recommended to
    allocate as large of a buffer as the trainer's CPU memory constraints will allow;
    note that the ceiling of the CPU memory usage on a given node will be
    ``(# of trainers on node) * max(memory used by prefetching, memory used by shuffle
    buffer)``, where the aggressiveness of the prefetching is controlled by the
    ``prefetch_blocks`` argument. See
    :meth:`ds.iter_batches() <ray.data.Dataset.iter_batches>` for details.

    .. literalinclude:: doc_code/air_ingest.py
        :language: python
        :start-after: __local_shuffling_start__
        :end-before: __local_shuffling_end__

.. tabbed:: Global Shuffling

    Global shuffling provides more uniformly random (decorrelated) samples and is carried
    out via a distributed map-reduce operation. This higher quality shuffle can often lead
    to more precision gain per training step, but it is also an expensive distributed
    operation and will decrease the ingest throughput. As long as the shuffled ingest
    throughput matches or exceeds the model training (forward pass, backward pass, gradient sync)
    throughput, this higher-quality shuffle shouldn't slow down the overall training.

    If global shuffling *is* causing the ingest throughput to become the training
    bottleneck, local shuffling may be a better option.

    .. literalinclude:: doc_code/air_ingest.py
        :language: python
        :start-after: __global_shuffling_start__
        :end-before: __global_shuffling_end__

Ingest and Ray Tune
-------------------

.. note::

    Train always uses Tune as the execution backend under the hood, even when running just ``Trainer.fit()`` directly (this is treated as a single-trial experiment). This ensures consistency of execution.

Placement Group Behavior
~~~~~~~~~~~~~~~~~~~~~~~~

Tune typically creates a placement group reserving resource for each of its trials. These placement groups only reserve resources for the Train actors, however. By default, Dataset preprocessing tasks run using "spare" CPU resources in the cluster, which enables better autoscaling and utilization of resources. It is also possible to set aside node CPUs for Dataset tasks using the ``_max_cpu_fraction_per_node`` option of DatasetConfig (Experimental).

.. warning::

    If trial placement groups reserve all the CPUs in the cluster, then it may be that no CPUs are left for Datasets to use, and trials can hang. This can easily happen when using CPU-only trainers. For example, if you can change the above ingest example to use ``ray.init(num_cpus=2)``, such a hang will happen.

Refer to the :ref:`Datasets in Tune Example <datasets_tune>` to understand how to configure your Trainer. We recommend starting with the default of allowing tasks to run using spare cluster resources, and only changing this if you encounter hangs or want more performance predictability.

Dataset Sharing
~~~~~~~~~~~~~~~

When you pass Datasets to a Tuner, it is important to understand that the Datasets are executed independently per-trial. This could potentially duplicate data reads in the cluster. To share Dataset blocks between trials, call `ds = ds.fully_executed()` prior to passing the Dataset to the Tuner. This ensures that the initial read operation will not be repeated per trial.

Debugging Ingest with the ``DummyTrainer``
------------------------------------------

Data ingest problems can be challenging to debug when combined in a full training pipeline. To isolate data
ingest issues from other possible training problems, we provide the ``ray.air.util.check_ingest.DummyTrainer``
utility class that can be used to debug ingest problems. Let's walk through using DummyTrainer to understand
and resolve an ingest misconfiguration.

Setting it up
~~~~~~~~~~~~~

First, let's create a synthetic in-memory dataset and setup a simple preprocessor pipeline. For this example,
we'll run it on a 3-node cluster with m5.4xlarge nodes. In practice we might want to use a single machine to
keep data local, but we'll use a cluster for illustrative purposes.

.. literalinclude:: doc_code/air_ingest.py
    :language: python
    :start-after: __check_ingest_1__
    :end-before: __check_ingest_1_end__

Next, we instantiate and fit a ``DummyTrainer`` with a single training worker and no GPUs. You can customize
these parameters to simulate your use training use cases (e.g., 16 trainers each with GPUs enabled).

.. literalinclude:: doc_code/air_ingest.py
    :language: python
    :start-after: __check_ingest_2__
    :end-before: __check_ingest_2_end__

Understanding the output
~~~~~~~~~~~~~~~~~~~~~~~~

Let's walk through the output. First, the job starts and executes preprocessing. You can see that the
preprocessing runs in ``6.8s`` below. The dataset stats for the preprocessing is also printed:

.. code::

    Starting dataset preprocessing
    Preprocessed datasets in 6.874227493000035 seconds
    Preprocessor Chain(preprocessors=(BatchMapper(fn=<lambda>), BatchMapper(fn=<lambda>)))
    Preprocessor transform stats:

    Stage 1 read->map_batches: 100/100 blocks executed in 4.57s
    * Remote wall time: 120.68ms min, 522.36ms max, 251.53ms mean, 25.15s total
    * Remote cpu time: 116.55ms min, 278.08ms max, 216.38ms mean, 21.64s total
    * Output num rows: 500 min, 500 max, 500 mean, 50000 total
    * Output size bytes: 102400128 min, 102400128 max, 102400128 mean, 10240012800 total
    * Tasks per node: 16 min, 48 max, 33 mean; 3 nodes used

    Stage 2 map_batches: 100/100 blocks executed in 2.22s
    * Remote wall time: 89.07ms min, 302.71ms max, 175.12ms mean, 17.51s total
    * Remote cpu time: 89.22ms min, 207.53ms max, 137.5ms mean, 13.75s total
    * Output num rows: 500 min, 500 max, 500 mean, 50000 total
    * Output size bytes: 102400128 min, 102400128 max, 102400128 mean, 10240012800 total
    * Tasks per node: 30 min, 37 max, 33 mean; 3 nodes used


When the train job finishes running, it will print out some more statistics.

.. code::

    P50/P95/Max batch delay (s) 1.101227020500005 1.120024863100042 1.9424749629999951
    Num epochs read 1
    Num batches read 100
    Num bytes read 9765.64 MiB
    Mean throughput 116.59 MiB/s

Let's break it down:

* **Batch delay**: Time the trainer spents waiting for the next data batch to be fetched. Ideally
  this value is as close to zero as possible. If it is too high, Ray may be spending too much time
  loading data from disk.
* **Num epochs read**: The number of times the trainer read the dataset during the run.
* **Num batches read**: The number of batches read.
* **Num bytes read**: The number of bytes read.
* **Mean throughput**: The average read throughput.

Finally, we can query memory statistics (this can be run in the middle of a job) to get an idea of how this workload used the object store.

.. code::

    ray memory --stats-only

As you can see, this run used 18GiB of object store memory, which was 32% of the total memory available
on the cluster. No disk spilling was reported:

.. code::

    --- Aggregate object store stats across all nodes ---
    Plasma memory usage 18554 MiB, 242 objects, 32.98% full, 0.17% needed
    Objects consumed by Ray tasks: 38965 MiB.

Debugging the performance problem
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

So why was the data ingest only 116MiB/s above? That's sufficient for many models, but one would expect
faster if the trainer was doing nothing except read the data. Based on the stats above, there was no object
spilling, but there was a high batch delay.

We can guess that perhaps AIR was spending too much time loading blocks from other machines, since
we were using a multi-node cluster. We can test this by setting ``prefetch_blocks=10`` to prefetch
blocks more aggressively and rerunning training.

.. code::

    P50/P95/Max batch delay (s) 0.0006792084998323844 0.0009853049503362856 0.12657493300002898
    Num epochs read 47
    Num batches read 4700
    Num bytes read 458984.95 MiB
    Mean throughput 15136.18 MiB/s

That's much better! Now we can see that our DummyTrainer is ingesting data at a rate of 15000MiB/s,
and was able to read through many more epochs of training. This high throughput means
that all data was able to be fit into memory on a single node.

Going from DummyTrainer to your real Trainer
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Once you're happy with the ingest performance of with DummyTrainer with synthetic data, the next steps are to switch to adapting it for your real workload scenario. This involves:

* **Scaling the DummyTrainer**: Change the scaling config of the DummyTrainer and cluster configuration to reflect your target workload.
* **Switching the Dataset**: Change the dataset from synthetic tensor data to reading your real dataset.
* **Switching the Trainer**: Swap the DummyTrainer with your real trainer.

Switching these components one by one allows performance problems to be easily isolated and reproduced.

Performance Tips
----------------

**Memory availability**: To maximize ingest performance, consider using machines with sufficient memory to fit the dataset entirely in memory. This avoids the need for disk spilling, streamed ingest, or fetching data across the network. As a rule of thumb, a Ray cluster with fewer but bigger nodes will outperform a Ray cluster with more smaller nodes due to better memory locality.

**Autoscaling**: We generally recommend first trying out AIR training with a fixed size cluster. This makes it easier to understand and debug issues. Autoscaling can be enabled after you are happy with performance to autoscale experiment sweeps with Tune, etc. We also recommend starting with autoscaling with a single node type. Autoscaling with hetereogeneous clusters can optimize costs, but may complicate performance debugging.

**Partitioning**: By default, Datasets will create up to 200 blocks per Dataset, or less if there are fewer base files for the Dataset. If you run into out-of-memory errors during preprocessing, consider increasing the number of blocks to reduce their size. To increase the max number of partitions, set the ``parallelism`` option when calling ``ray.data.read_*()``. To change the number of partitions at runtime, use ``ds.repartition(N)``. As a rule of thumb, blocks should be not more than 1-2GiB each.