GCP Dataflow

Apache Beam Pipelines on Cloud Dataflow

Pipeline authoring

Use this section when implementing Dataflow pipeline logic using Apache Beam.

Check if existing Google Dataflow Template exists

Google provides a variety of pre-built, open source Dataflow templates that can be used for common scenarios. Before implementing a pipeline from scratch, you MUST follow the steps below to check whether a Dataflow template for the pipeline logic you need to implement already exists.

• Step 1: Check for a matching Google Dataflow Template

  • Identify the source and sink (e.g., GCS to BigQuery) from the user's request. Note: You MUST NOT proceed until the source and sink are clearly identified.
  • Action: List templates in the public dataflow-templates bucket:
    • For Classic templates, check gs://dataflow-templates/latest.
    • For Flex templates, check gs://dataflow-templates/latest/flex. Use gcloud storage ls to list the contents.
  • Match templates by name or description to the source and sink.
  • If no matching template is found, go to Create a new pipeline from scratch.

• Step 2: Confirm template selection

  • Present the matched template(s) to the user with a brief explanation of why they match, and make a note of whether it is a Classic or Flex template.
  • Action: Ask the user for explicit confirmation to proceed with this template.
  • If the user rejects or prefers a custom solution, proceed to Create a new pipeline from scratch.

Create a new pipeline from scratch

Use this section when creating a new project for a Dataflow pipeline from scratch.

• If the user doesn't say explicitly which language (Java, Python, Go) shall be used to write the pipeline, you MUST confirm the language.
• Determine which version of Beam SDK should be used by searching for the most recently released version of Apache Beam, unless the user already uses a particular version.
  • Action: Run a web search for the latest Apache Beam SDK release.
• YOU MUST use same version of Apache Beam consistently throughout the project in Dockerfiles, requirements.txt, and other similar files where versions are specified.

Java projects using Gradle

Use this section when configuring a Dataflow Java pipeline project using gradle.

• Shadow Jars (Fat Jars): Do NOT propose to use the Shadow plugin (com.github.johnrengelman.shadow) unless the user explicitly requests a Fat Jar.
• Passing command-line parameters: Use the application plugin for passing command-line parameters.
• SLF4J Logging Dependency Alignment:
  • Verify the slf4j-api version pulled transitively by Apache Beam.
  • You MUST configure the application logging backend (slf4j-simple, logback-classic, etc.) to exactly match the major/minor version of the resolved slf4j-api.

Packaging a pipeline as a Flex Template

Use this section to package pipeline code as a Flex template.

Flex Templates offer a hermetic and reproducible launch environment for a pipeline. They are easy to launch with gcloud or with orchestrators like Cloud Composer. You MUST package the pipeline as a Flex Template when creating new Dataflow pipeline projects.

Follow the steps below:

• Provide Instructions: Provide instructions on rebuilding and running Flex Templates to the user in walkthrough.
• Use Single Docker Image for Python pipelines: For Python Flex Templates, it is better to use a single image for the template launcher image and for the worker runtime environment (--sdk_container_image). Does the Python pipeline require extra dependencies (e.g., using --requirements_file, --setup_file, or --extra_package)? If so, YOU MUST recommend the Single Docker Image Configuration for the Flex Template. See python_flex_template_reference.md for details.
• Prefer Cloud Build over Local Docker:
  • Do NOT assume local Docker availability on the workspace machine.
  • Action: Suggest and provide cloudbuild.yaml out-of-the-box for building and pushing images unless local setup is explicitly requested.
  • When building images with Cloud Build in the background you MUST provide the link where the user can monitor the long-running operation.
• Providing SSL certificates and Secrets to Workers:
  • If certificates or keys are stored in Secret Manager, NEVER bake them into the Docker image layers. Instead, retrieve them dynamically at runtime inside the Apache Beam DoFn.setup() lifecycle using the Secret Manager client library (writing them to ephemeral worker disk like /tmp only if physical file paths are strictly required). Ensure the Dataflow Worker Service Account has the roles/secretmanager.secretAccessor role.

Configuring Google-provided templates

Use this section when the user has selected a Google-provided template (Classic or Flex) and you need to configure it.

• Step 1: Get template metadata

  • Identify template type:
    • Classic: Metadata files are in gs://dataflow-templates and end with _metadata (e.g., gs://dataflow-templates/latest/Word_Count_metadata).
    • Flex: Metadata are embedded in the template spec file under gs://dataflow-templates/latest/flex (e.g. gs://dataflow-templates/latest/flex/Cloud_Datastream_to_BigQuery).
  • Read the corresponding template metadata file to identify required parameters.
  • Note:
    • Make sure to run a recursive search over the bucket if needed to locate the metadata.
    • If the template parameters include UDF-related fields (e.g., javascriptTextTransformGcsPath, javascriptTextTransformFunctionName), refer to the UDF guide to write and configure the UDF.
    • Parameter-Based SSL / Secret Staging: If the Google-provided template requires local SSL certificates or Secret Manager secrets, pass comma-separated GCS paths via the extraFilesToStage parameter. The runner will drop them into /extra_files on worker VMs. Refer to the SSL certificates guide for local referencing syntax (/extra_files/...).

• Step 2: Get network configuration

  • Action: Run gcloud commands to list networks and subnetworks.
  • Confirm the network and subnetwork to use with the user.

• Step 3: Identify required parameters and prepare resources

  • Extract required parameters from the template metadata.

  • [!IMPORTANT]

  • Strict parameter validation: Any parameter in the metadata JSON

  • that does NOT explicitly have "isOptional": true is **strictly

  • required** by the Dataflow API.

  • This applies even if the description suggests it has a default value

  • (e.g., csvFormat or badRecordsOutputTable in some templates).

  • You must identify and supply all of them.

  • Identify which parameters are provided by the user and which need to be resolved or created by you.
    • Action: Present these parameters to the user using Markdown Key-Value (bullet points) for clarity and confirmation.
  • Schema Handling: If a schema JSON parameter (like schemaJSONPath or JSONPath) is required:
    • Action: Ask the user to provide the GCS path to an existing schema file or the JSON content.
    • If the user does not have a schema file, ask them to provide the field names and their types. Construct the schema JSON locally and present it to the user for validation.
    • Once confirmed by the user, write the schema JSON file locally and upload it to a GCS staging location, then supply this path to the parameter.
  • Pre-create Target Sink (Best Practice): To ensure stability and avoid runtime creation schema mismatches:
    • Action: Clarify with the user whether the target sink (e.g., BigQuery table, Spanner database/table) already exists. If the user confirms it exists, proceed to the remaining steps as-is.
    • If it does not exist, ask for permission to create it. If permitted, create it yourself. Otherwise, provide the exact creation commands to the user.
    • If the sink is BigQuery, refer to Destination-specific prerequisites for crucial table and error table setup.
  • Include additional parameters such as service accounts, network details, and other pipeline options.
    • Specifying Options: For Google-provided Flex Templates, refer to the Specifying options for Flex Templates guide for how to pass parameters and additional experiments.

Destination-specific prerequisites

Different templates might require specific resources to be prepared in the target sink before execution. Follow the instructions for your target sink below.

BigQuery

When running templates that write to BigQuery, you MUST ensure the following resources are prepared to prevent job failures:

• Pre-create Target Table: Create the target BigQuery table (e.g., using bq mk) before launching the job. Ensure the schema matches the template's expectations.
• Pre-create Error/Bad Records Table: Many templates that write to BigQuery have a parameter for redirecting failed records (e.g., badRecordsOutputTable or outputDeadletterTable). Some templates attempt to auto-create this table. However, pre-creating it is a best practice. This ensures correct schema and permissions.
  • How to Determine the Error Schema: Trace the schema definition in the public DataflowTemplates GitHub repository:
    1. Locate the source code or README for the template you are using (e.g., in v1/ or v2/ directories).
    2. Identify the parameter name used for the error table (e.g., badRecordsOutputTable or outputDeadletterTable).
    3. Search the source code to see how the schema is defined or loaded for that parameter:
       • Example (Code Reference): In PubSubToBigQuery.java, the schema is set using ResourceUtils.getDeadletterTableSchemaJson(). Tracing ResourceUtils.java shows it loads the schema from streaming_source_deadletter_table_schema.json on GitHub.
       • Example (Documentation Reference): For simpler templates, the schema might be listed in the official documentation, such as the RawContent/ErrorMsg schema shown in the CSV to BigQuery DevSite Doc.

Configuring Custom Pipelines (Dataflow Runner)

Use this section when preparing to run a custom Apache Beam pipeline on Dataflow.

• When launching Python Pipelines without a Flex Template with DataflowRunner, you MUST scan the pipeline project directory for the following files:

  • requirements.txt:
    • If found, you MUST include --requirements_file pipeline option.
  • setup.py:
    • If found, you MUST include --setup_file pipeline option. This is critical if the pipeline uses local modules or packages.

• When launching Python Pipelines with a Flex Template, if the Flex Template image is also the SDK Container image (Single Docker Image Configuration), then you MUST supply the image in the sdk_container_image parameter.

Lookup environment resources instead of using placeholder values

• Avoid using generic placeholders (e.g., your-gcp-project-id) for GCP resources when drafting run scripts or configs. Action: If values are unknown, proactively run commands like gcloud config get-value project to find active resources to pre-fill scripts for the user. Confirm the values with the user before proceeding.

Job Execution

Use this section when configuration is complete and you are ready to launch any Dataflow job (Google-provided template, Custom Flex template, or standalone pipeline).

Universal Execution Workflow

1. Construct Launch Command: Draft the full launch command based on the pipeline type (e.g., gcloud dataflow flex-template run or python main.py --runner=DataflowRunner). Ensure workers default to private IP configuration unless specified otherwise, and verify target project permissions.
2. Mandatory Pre-Launch Confirmation: Present the entire drafted command to the user at once. Explain the purpose of all parameters (including experimental flags) and allow the user to review and correct the command as a batch instead of confirming piecemeal. Do NOT proceed with execution until explicitly approved.
3. Trigger Job: Once approved, execute the command and note the resulting Job ID (displaying it to the user).
4. Display Console URL: Construct and present the direct Cloud Console monitoring URL: https://console.cloud.google.com/dataflow/jobs//?project=

Job Monitoring

Use this section to monitor the progress of a running Dataflow job.

• Check the status of the triggered Dataflow job using the job ID.
• Run the check every 30 seconds for the first 2 minutes, then check every 3 minutes, unless specified otherwise by the user.
• Note: Do NOT perform data check queries on the sink until the job has reached a stable RUNNING or DONE state.

Diagnostics & Troubleshooting

[!IMPORTANT] YOU MUST use this section when the user asks about performance of their Dataflow pipelines. This can be used to debug issues like pipeline slowness, pipeline failures, etc.

Task Execution Workflow

1. Understand User Request: Extract Job ID, Project ID, Transform Name (optional), and Time Window.

2. Transform Name Mapping: If the user requires transform-based debugging, map user-provided Transform Names to actual Dataflow stage or ptransform and apply to filters while querying:

   This mapping can be extracted from gcloud dataflow jobs describe JOB_ID --full --format="json(pipelineDescription.executionPipelineStage)".

   1. Extract the targets:
      • Get stage_id: name property at the parent stage level. This matches "F[digit]" (e.g. "F6").
      • Get ptransform: inside the componentTransform array, read precisely from userName or originalTransform (e.g. "RateLimitAndLog/ParMultiDo(RateLimitAndLog)"). and use it as ptransform.
   2. Apply the filters strictly following mapping mechanics:
      • For Cloud Logging queries: Apply extracted ptransform name to filter resource.labels.step_id="[Extracted ptransform name]".
      • For Monitoring queries: Use the stage_id/ptransform filters based on filters supported by metric: metric.labels.ptransform="[Extracted ptransform name]" or metric.labels.stage="[Extracted stage_id]".

3. Query Telemetry:

   • Use Dataflow REST API to get High level Job Messages/Events that happened in the job.
   • Refer to dataflow_diagnostics_reference.md for key metrics and logging query patterns based on Job Type.
   • Use Monitoring REST API to fetch metrics.
   • Use GCloud Logging command to fetch logs.
   • Use Dataflow REST API to fetch current snapshot metrics when historical time-series are not needed.

4. Analysis:

   • For Streaming Jobs
     • Overall Job Health: YOU MUST refer to streaming_job_health to analyze overall streaming job health.
     • Analyze Bottlenecks and Parallelism. YOU MUST refer to bottlenecks_and_parallelism_context and interpret the bottlenecks and parallelism metrics in that context.
     • Analyze Autoscaling Behavior. YOU MUST refer to streaming_horizontal_autoscaling_analysis.md
   • For Batch Jobs
     • Correlate metrics spikes/drops with log errors.
     • Identify Issues.

5. Output: Provide a synthesized diagnosis containing symptoms, root causes, and target code links (using file:///... format). Strictly follow the response structure appropriate for the job type:

   For Streaming Jobs:

   1. Overall Job State: State categorization (Healthy, Mostly Healthy, Not Healthy) per streaming_job_health.
   2. High-level Job Events: Notable control plane events, errors, or stage failures parsed from job messages.
   3. Data Freshness: Current data delay utilizing job/data_watermark_age / job/per_stage_data_watermark_age and system lag.
   4. Throughput: Processing rate trends utilizing job/elements_produced_count / job/estimated_bytes_produced_count.
   5. Backlog: Input backlog (if source stage) or inter-stage backlog using job/estimated_backlog_processing_time / job/backlog_bytes.
   6. Bottlenecks & Parallelism: Queue delay diagnostics using job/is_bottleneck (interpreting likely_cause / bottleneck_kind) and key metrics job/backlogged_keys / job/processing_parallelism_keys interpreted in the context of bottlenecks_and_parallelism_context.
   7. Autoscaling Analysis: Scaling trends using job/horizontal_worker_scaling (and label rationale), clamp limits (job/max_worker_instances_limit / job/min_worker_instances_limit), and utilization hints in the context of streaming_horizontal_autoscaling_analysis.
   8. Recommendations: Direct remediation plans (in-flight updates, client-side configurations, or code corrections linked via absolute file:/// URIs).

   For Batch Jobs:

   1. High-level Job Events: Notable control plane events, errors, or stage failures parsed from job messages.
   2. Throughput: Processing rate trends utilizing job/elements_produced_count (primary performance indicator).
   3. Recommendations: Direct remediation plans to future runs (client-side configurations, or code corrections linked via absolute file:/// URIs).