connectors

No menu items for this category

Best Practices for Metadata Ingestion

In this section we are going to present some guidelines that can be useful when preparing metadata ingestion both from the UI or via any custom orchestration system.

We will use the generic terms from Airflow, as the most common used tool, but the underlying ideas can be applied anywhere.

  • DAGs should not have any retry: If the workflow is marked as failed due to any error (unexpected exception, connectivity issues, individual assets’ errors,...) there is usually no point on running automatic retries. For heavy workflows failing in the middle of processing, it will just incur in extra costs.

    Note that for internal communications between the Ingestion Workflow and the OpenMetadata APIs, we already have an internal retry in place in case of intermittent networking issues.

  • DAGs should not have a catch-up: Any ingestion will be based on the current state of data and metadata. If old runs were skipped for any reason, there is no point in triggering past executions as they won’t be adding any value. Just the single, most recent run will already be providing all the information available.

  • Be mindful of enabled DEBUG logs: When configuring the ingestion YAML you have the option to control the logging level. Keeping it as INFO (default) is the usual best bet. Only use DEBUG logs when testing out ingestion for the first time

  • Test the ingestions using the CLI if you will be building a DAG: When preparing the first ingestion processes, it is ok to try different configurations (debug logs, enable views, filtering of assets,...). The fastest and easiest way to test the ingestion process that will end up on a DAG is using the CLI (example). Playing with the CLI will help you find the right YAML configuration fast. Note that for OpenMetadata, the process that gets triggered from the CLI, is the same as the one that will eventually run in your DAGs. If you have the possibility to test the CLI first, it will give you fast feedback and will help you isolate your tests.

  • Apply the right filters: For example, there is usually no business-related information on schemas such as INFORMATION_SCHEMA. You can use OpenMetadata filtering logic on databases, services and tables to opt in/out specific assets.
  • On filters, scheduling and asset importance: While OpenMetadata provides sampling and multi-threading, profiling can be a costly and time-consuming process. Then it is important to know which data assets are business critical.
    • Deploy multiple profiler ingestions for the same service: For a given service, prepare different ingestion pipelines, each of them attacking a specific set of assets based on input filters. You can then schedule more important assets to be profiled more often, while keeping the rest of profiles to be executed either on demand, or with lower cadence.
  • Apply the right sampling: Important tables can hold higher sampling, while the rest of assets might be good enough with smaller %.
  • Schedule and log duration should match: The Log Duration configuration parameter specifies how many days in the past we are going to look for query history data. If we schedule the workflows to run daily, there is no need to look for the past week, as we will be re-analysing data that won’t change.

OpenMetadata Ingestion Troubleshooting

Here we will discuss different errors that you might encounter when running a workflow:

  • Connection errors: When deploying ingestions from the OpenMetadata UI you have the possibility to test the connection when configuring the service. This connection test happens at the Airflow host configured with OpenMetadata. If instead, you are running your ingestion workflows from any external system, you’ll need to validate that the host where the ingestion runs has the proper network settings to reach both the source system and OpenMetadata.
  • Processing Errors: During the workflow process you might see logs like Cannot ingest X due to Y or similar statements. They appear for specific assets being ingested, and the origin can be different:
    • Missing permissions on a specific table or tag (e.g., due to BigQuery policies),
    • Internal errors when processing specific assets or translating them to the OpenMetadata standard. In these cases, you can reach out to the OpenMetadata team. The workflow itself will continue, and the OpenMetadata team can help analyse the root cause and provide a fix.
  • Workflow breaking exceptions: In rare circumstances there can be exceptions that break the overall workflow processing. The goal of the Ingestion Framework is to be as robust as possible and continue even for specific assets failures (see point above). If there is a scenario not contemplated by the current code, the OpenMetadata team will apply the highest priority to fix the issue and allow the workflow to run end to end.