-
Notifications
You must be signed in to change notification settings - Fork 123
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Cache TaskGroup/DAG regardless of the load_method
#926
Comments
…ct (#904) Improve the performance to run the benchmark DAG with 100 tasks by 34% and the benchmark DAG with 10 tasks by 22%, by persisting the dbt partial parse artifact in Airflow nodes. This performance can be even higher in the case of dbt projects that take more time to be parsed. With the introduction of #800, Cosmos supports using dbt partial parsing files. This feature has led to a substantial performance improvement, particularly for large dbt projects, both during Airflow DAG parsing (using LoadMode.DBT_LS) and also Airflow task execution (when using `ExecutionMode.LOCAL` and `ExecutionMode.VIRTUALENV`). There were two limitations with the initial support to partial parsing, which the current PR aims to address: 1. DAGs using Cosmos `ProfileMapping` classes could not leverage this feature. This is because the partial parsing relies on profile files not changing, and by default, Cosmos would mock the dbt profile in several parts of the code. The consequence is that users trying Cosmos 1.4.0a1 will see the following message: ``` 13:33:16 Unable to do partial parsing because profile has changed 13:33:16 Unable to do partial parsing because env vars used in profiles.yml have changed ``` 2. The user had to explicitly provide a `partial_parse.msgpack` file in the original project folder for their Airflow deployment - and if, for any reason, this became outdated, the user would not leverage the partial parsing feature. Since Cosmos runs dbt tasks from within a temporary directory, the partial parse would be stale for some users, it would be updated in the temporary directory, but the next time the task was run, Cosmos/dbt would not leverage the recently updated `partial_parse.msgpack` file. The current PR addresses these two issues respectfully by: 1. Allowing users that want to leverage Cosmos `ProfileMapping` and partial parsing to use `RenderConfig(enable_mock_profile=False)` 2. Introducing a Cosmos cache directory where we are persisting partial parsing files. This feature is enabled by default, but users can opt out by setting the Airflow configuration `[cosmos][enable_cache] = False` (exporting the environment variable `AIRFLOW__COSMOS__ENABLE_CACHE=0`). Users can also define the temporary directory used to store these files using the `[cosmos][cache_dir]` Airflow configuration. By default, Cosmos will create and use a folder `cosmos` inside the system's temporary directory: https://docs.python.org/3/library/tempfile.html#tempfile.gettempdir . This PR affects both DAG parsing and task execution. Although it does not introduce an optimisation per se, it makes the partial parse feature implemented #800 available to more users. Closes: #722 I updated the documentation in the PR: #898 Some future steps related to optimization associated to caching to be addressed in separate PRs: i. Change how we create mocked profiles, to create the file itself in the same way, referencing an environment variable with the same name - and only changing the value of the environment variable (#924) ii. Extend caching to the `profiles.yml` created by Cosmos in the newly introduced `tmp/cosmos` without the need to recreate it every time (#925). iii. Extend caching to the Airflow DAG/Task group as a pickle file - this approach is more generic and would work for every type of DAG parsing and executor. (#926) iv. Support persisting/fetching the cache from remote storage so we don't have to replicate it for every Airflow scheduler and worker node. (#927) v. Cache dbt deps lock file/avoid installing dbt steps every time. We can leverage `package-lock.yml` introduced in dbt t 1.7 (https://docs.getdbt.com/reference/commands/deps#predictable-package-installs), but ideally, we'd have a strategy to support older versions of dbt as well. (#930) vi. Support caching `partial_parse.msgpack` even when vars change: https://medium.com/@sebastian.daum89/how-to-speed-up-single-dbt-invocations-when-using-changing-dbt-variables-b9d91ce3fb0d vii. Support partial parsing in Docker and Kubernetes Cosmos executors (#929) viii. Centralise all the Airflow-based config into Cosmos settings.py & create a dedicated docs page containing information about these (#928) **How to validate this change** Run the performance benchmark against this and the `main` branch, checking the value of `/tmp/performance_results.txt`. Example of commands run locally: ``` # Setup AIRFLOW_HOME=`pwd` AIRFLOW_CONN_AIRFLOW_DB="postgres://postgres:postgres@0.0.0.0:5432/postgres" PYTHONPATH=`pwd` AIRFLOW_HOME=`pwd` AIRFLOW__CORE__DAGBAG_IMPORT_TIMEOUT=20000 AIRFLOW__CORE__DAG_FILE_PROCESSOR_TIMEOUT=20000 hatch run tests.py3.11-2.7:test-performance-setup # Run test for 100 dbt models per DAG: MODEL_COUNT=100 AIRFLOW_HOME=`pwd` AIRFLOW_CONN_AIRFLOW_DB="postgres://postgres:postgres@0.0.0.0:5432/postgres" PYTHONPATH=`pwd` AIRFLOW_HOME=`pwd` AIRFLOW__CORE__DAGBAG_IMPORT_TIMEOUT=20000 AIRFLOW__CORE__DAG_FILE_PROCESSOR_TIMEOUT=20000 hatch run tests.py3.11-2.7:test-performance ``` An example of output when running 100 with the main branch: ``` NUM_MODELS=100 TIME=114.18614888191223 MODELS_PER_SECOND=0.8757629623135543 DBT_VERSION=1.7.13 ``` And with the current PR: ``` NUM_MODELS=100 TIME=75.17766404151917 MODELS_PER_SECOND=1.33018232576064 DBT_VERSION=1.7.13 ```
load_method
of choiceload_method
of choice
load_method
of choiceload_method
load_method
load_method
Progress can be seen in draft PR: #992 |
I explored and validated a few possible ways of quickly generating the "version" of the cache. Some are safer than others, the idea was to get a sense of cost/benefit:
So far, it seems |
I extended the initial implementation to: I also updated the PR description with what is missing:
|
When using the current approach in a distributed environment, there are two challenges:
We'll look into improving this. Examples of the behaviour in a distributed Airflow environment: |
Context
PR #904 introduced the first caching implementation in Cosmos on top of the previously implemented support for partial parsing (#800).
Although these changes contributed significantly to the performance of Cosmos, this approach does not improve the DAG parsing performance for methods other than
dbt ls
. In those cases, Cosmos will still:During DAG parsing - which happens at the Airflow DAG Processor and is executed every time a task is scheduled/run.
This task aims to extend the cache support to allow the cache of the complete DAG/TaskGroup, without the need to do the two previously described steps.
Acceptance criteria
dbt
project or to the Airflow DAGload_method=LoadMode.DBT_MANIFEST
The text was updated successfully, but these errors were encountered: