add airflow example

ghstack-source-id: 571c14f
Pull Request resolved: #33

Author: hudeven

torchrecipes/paved_path/README.md

@@ -58,3 +58,6 @@ Note that it will output a URL like "aws_batch://torchx/..." that is used to tra
 torchx status "aws_batch://torchx/..."
 ```
 
+## Pipelines
+As your applications getting complicated, you can make them as pipelines, manage and monitor them by frameworks like Airflow, Kubeflow, etc.
+* [Airflow example](https://github.com/facebookresearch/recipes/tree/main/torchrecipes/paved_path/airflow)

torchrecipes/paved_path/airflow/README.md

@@ -0,0 +1,22 @@
+# Airflow example
+1. Install and start an Airflow server
+```bash
+./setup.sh
+```
+> **_NOTE:_**: Airflow UI can be accessed at http://0.0.0.0:8080 (replace the address with your EC2 instance address for public access). Learn more about airflow from [Quick Start](https://airflow.apache.org/docs/apache-airflow/stable/start/local.html)
+
+2. Create a dag config. See the example in `train_charnn.py`
+
+3. Set `dag_folder` to folder containing the dag config in `~/airflow/airflow.cfg`. Such that Airflow can discover your dag configs.
+
+3. Run a task instance
+```bash
+airflow tasks run train_charnn train 2022-08-01
+```
+> **_NOTE:_**: the instance can be monitored in UI: http://0.0.0.0:8080/taskinstance/list
+
+4. Backfill the dag over 2 days
+```bash
+airflow dags backfill train_charnn --start-date 2022-08-01 --end-date 2022-08-02
+```
+> **_NOTE:_**: the dag runs can be monitored in UI: http://0.0.0.0:8080/dagrun/list/

torchrecipes/paved_path/airflow/setup.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+install_airflow=true
+start_local_airflow=true
+
+if [ "$install_airflow" = true ]
+then
+    pip3 install --upgrade pip
+    sudo apt install libffi-dev
+    pip3 install setuptools-rust
+    pip3 install "apache-airflow[celery]==2.3.0" --constraint "https://raw.githubusercontent.com/apache/airflow/constraints-2.3.0/constraints-3.8.txt"
+    pip3 install apache-airflow-providers-amazon
+    pip3 install boto3
+fi
+
+# https://airflow.apache.org/docs/apache-airflow/stable/start/local.html
+if [ "$start_local_airflow" = true ]
+then
+    airflow standalone
+fi

torchrecipes/paved_path/airflow/train_charnn.py

@@ -0,0 +1,98 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+from datetime import datetime, timedelta
+
+import boto3.session
+
+from airflow import DAG
+from airflow.operators.bash import BashOperator
+from airflow.operators.python import PythonOperator
+from airflow.providers.amazon.aws.hooks.batch_waiters import BatchWaitersHook
+
+REGION = "us-west-2"
+JOB_QUEUE = "torchx-gpu"
+ECR_URL = os.environ["ECR_URL"]
+
+
+default_args = {
+    "depends_on_past": False,
+    "email": ["[email protected]"],
+    "email_on_failure": False,
+    "email_on_retry": False,
+    "retries": 0,
+    "retry_delay": timedelta(minutes=5),
+}
+
+
+dag = DAG(
+    "train_charnn",
+    default_args=default_args,
+    description="A DAG to train charnn in AWS Batch",
+    schedule_interval="@daily",
+    catchup=False,
+    start_date=datetime(2022, 8, 1),
+    tags=["aws_batch"],
+)
+
+
+# This example uses torchx CLI with BashOperator.
+# We can also use PythonOperator to achive it.
+train = BashOperator(
+    task_id="train",
+    bash_command=f"""AWS_DEFAULT_REGION=$REGION \
+        torchx run --workspace '' -s aws_batch \
+        -cfg queue={JOB_QUEUE},image_repo={ECR_URL}/charnn dist.ddp \
+        --script charnn/main.py --image {ECR_URL}/charnn:latest \
+        --cpu 8 --gpu 2 -j 1x2 --memMB 20480 2>&1 \
+        | grep -Eo aws_batch://torchx/{JOB_QUEUE}:main-[a-z0-9]+""",
+    env={
+        "REGION": REGION,
+        "JOB_QUEUE": JOB_QUEUE,
+        "ECR_URL": ECR_URL,
+    },
+    append_env=True,
+    dag=dag,
+    do_xcom_push=True,
+)
+
+
+def wait_for_batch_job(**context) -> bool:
+    session = boto3.session.Session()
+    client = session.client("batch", region_name=REGION)
+    # XComs are a mechanism that let Tasks talk to each other
+    # Learn more from https://airflow.apache.org/docs/apache-airflow/stable/concepts/xcoms.html
+    job = context["ti"].xcom_pull(task_ids="train")
+    job_desc = job.split("/")[-1]
+    queue_name, job_name = job_desc.split(":")
+    job_id = client.list_jobs(
+        jobQueue=queue_name,
+        filters=[{"name": "JOB_NAME", "values": [job_name]}],
+    )["jobSummaryList"][0]["jobId"]
+    waiter = BatchWaitersHook(region_name=REGION)
+    try:
+        waiter.wait_for_job(job_id)
+        return True
+    except Exception:
+        return False
+
+
+wait_for_job = PythonOperator(
+    task_id="wait_for_job",
+    python_callable=wait_for_batch_job,
+    dag=dag,
+)
+
+
+parse_output = BashOperator(
+    task_id="parse_output",
+    bash_command="echo {{ ti.xcom_pull(task_ids='wait_for_job') }}",
+    dag=dag,
+)
+
+
+train >> wait_for_job >> parse_output