Skip to content

Test Requirements

This document specifies the testing required for GCP Ingestion components.

Exceptions

Code that does not comply with this standard before it is deployed to production must include a document explaining why that decision was made.

Test Phases

  1. Continuous Integration
  2. Must run Unit Tests and Integration Tests
  3. Must run against every merge to main and every pull request
  4. Must block merging to main
  5. Pre-Production
  6. Must run Unit Tests and Load Tests
  7. Must run against every version deployed to production
  8. Must mimic production configuration as closely as possible
  9. Should block deployment to production
  10. Async
  11. Must run Slow Load Tests
  12. Must run against every version deployed to production
  13. Must mimic production configuration as closely as possible
  14. Should not block deployment to production
  15. Must notify someone to investigate failures

Test Categories

Tests must be both thorough and fast enough to not block development. They are split into categories with increasing run time and decreasing coverage.

  1. Unit Tests
  2. Must cover 100% of code behavior
  3. Should run as fast as possible
  4. Integration Tests
  5. Must cover all expected production behavior
  6. Must run fast enough to allow frequent merging to main
  7. Load Tests
  8. Must cover performance at scale with and without external services down
  9. Must run fast enough to allow multiple production deployments per day
  10. Slow Load Tests
  11. Must cover performance at scale with extended downtime
  12. May take a very long time

Unit Tests

  • Must run in under 5 seconds in CI, not including build time
  • May use CI parallelism and run each group of tests in under 5 seconds in CI
  • Should validate code as quickly as possible
  • Must run completely in memory
  • Must safely run in parallel
  • May mock anything that would not run in memory
  • Should cover 100% of branches and statements
  • Should cover all variations of statements having variable behavior without branching
  • Should cover all production configuration variations
  • Must have Integration Tests to cover exceptions
  • Must cover all input and output variations
  • Must cover all response codes
  • Must cover all valid URL patterns
  • Must cover gzipped and not gzipped
  • Must cover present, absent, and invalid required attributes
  • Must cover present, absent, and invalid optional attributes
  • Must cover present, absent, invalid, and wrong type payload JSON
  • Must cover present, absent, invalid, and wrong type payload fields
  • Must cover document_id duplicates
  • Must cover all error types
  • Must cover PubSub returning OK but rejecting messages in a batch
  • Must cover External service returns server error
  • Must cover External service timeout
  • Must cover External resource missing
  • Must cover Insufficient permissions
  • Must cover behavior when disk becomes full
  • Must cover behavior with disk already full
  • Must cover behavior when ingestion-edge disk is no longer full

Integration Tests

  • Must run in under 5 minutes in CI, not including build time
  • Must be configurable to run locally and in CI
  • Must be configurable to run against a deployment in GCP
  • May require a proxy to modify cloud service behavior
  • May require specific configuration
  • May skip coverage that cannot be forcibly produced in Dataflow
  • Must clean up created resources when complete
  • Must cover everything unit tests should but do not
  • Must cover a configuration that mimics production as closely as possible
  • Must cover all code with input from or output to external services
  • The ingestion-edge disk queue is an external service
  • Must cover all production configuration options
  • Must cover all API endpoints with invalid traffic
  • Must cover all API endpoints with healthy external services
  • Must assert disk queue is not used in ingestion-edge
  • Must assert inputs and outputs preserve schema in ingestion-beam
  • Must cover behavior with errors from external services
  • Must cover PubSub returning OK but rejecting messages in a batch
  • Must cover external service DNS does not resolve
  • Must cover external service cannot TCP connect
  • Must cover external service hangs after TCP connect
  • Must cover external service returns 500
  • Must cover resource does not exist in external service
  • Must cover insufficient permissions in external service
  • Must cover behavior when disk becomes full
  • Must cover behavior with disk already full
  • Must assert automatic recovery unless manual intervention is required
  • Must cover behavior when ingestion-edge disk is no longer full

Load Tests

  • Must run in parallel in under 1 hour
  • May require a separate deployment for each test
  • Must record how much it cost to run
  • Must not require manual intervention to evaluate
  • Must check performance against a tunable threshold
  • Must sustain at least 1.5x expected peak production traffic volume
  • As of 2018-08-01 peak production traffic volume is about 11K req/s
  • Must cover behavior before, during, and after a period of:
  • Service downtime, except ingestion-edge
  • 100%, 50%, and 5% invalid traffic
  • External services cannot TCP connect
  • 100% and 10% of external service requests time out
  • 100% and 10% external service requests return 500
  • Must cover ingestion-edge behavior with PubSub returning 500 and auto scaling due to disk filling up
  • Must assert ingestion-edge only writes to disk queue when PubSub is unhealthy
  • Must assert full recovery from downtime does not take longer than the downtime period

Slow Load Tests

  • May run for longer than 1 hour
  • Should not block deployment to production
  • Must cover behavior with sustained load and a backlog equivalent to:
  • For ingestion-edge: 1.5x the longest PubSub outage in the last year
    • As of 2018-08-01 the longest PubSub outage was about 4.5 hours
  • For everything else: 4 days (1 holiday weekend)
    • Should reach full recovery in under 1 day
  • Must cover behavior of ingestion-beam BigQuery output with sustained load and BigQuery returning 500 for at least 1.5x the longest BigQuery outage in the last year
  • As of 2018-08-01 the longest BigQuery outage was about 2.3 hours