Real-Time Analytics Platform  ·  View 17 of 21  ·  Operations

CI/CD and Environment Promotion

How a change reaches production, what stops a bad one at each gate, and how a streaming job is released without skipping an event.

Editable source SVG draw.io All views
Source
Source
Terraform Modules
infrastructure as code
Terraform Modules...
Azure Repos
trunk + short branches
Azure Repos...
Schema Contracts
versioned Avro
Schema Contracts...
Build and test
Build and test
Unit & PySpark Tests
Unit & PySpark Tests
Schema Compat Gate
Schema Compat Gate
SAST & IaC Scan
Defender for DevOps
SAST & IaC Scan...
Dev
Dev
Ephemeral ADX Database
torn down nightly
Ephemeral ADX Database...
Dev Stream Job
feature branch
Dev Stream Job...
Synthetic Event Generator
replayed production shape
Synthetic Event Generator...
Test and UAT
Test and UAT
Shadow Traffic
5% mirrored
Shadow Traffic...
Latency Gate
5 s p95 end to end
Latency Gate...
Consumer Contract Tests
Consumer Contract Tests
Production
Production
Change Approval
Change Approval
Blue/Green Stream Job
new consumer group
Blue/Green Stream Job...
View Rebuild Guard
no read downtime
View Rebuild Guard...
Post-release
Post-release
Error Budget Update
Error Budget Update
48 h Watch
lag and freshness
48 h Watch...
Auto Rollback
on lag breach
Auto Rollback...
breaking
breaking
revert
revert
CI/CD and Environment Promotion
CI/CD and Environment Promotion
Security / platform
Security / platform
Application we own
Application we own
Decision point
Decision point
Data store
Data store
failure / alternate
failure / alternate
Blue/green uses a second consumer group replaying from the same offsets, so no event is skipped at cut-over.
Blue/green uses a second consumer group replaying from the same offsets, so no event is skipped at cut-over.
v 1.0 · owner Platform Engineering · date 2026-08
v 1.0 · owner Platform Engineering · date 2026-08
Text is not SVG - cannot display

The gates that matter

  • Schema compatibility gate — a breaking contract change cannot merge
  • Latency gate — a build that pushes end-to-end p95 past 5 seconds does not promote
  • Consumer contract tests run against the actual registered schema, not a fixture

Zero-downtime release

  • Blue/green uses a second consumer group starting from the same committed offsets
  • Green is verified against shadow output before blue is stopped, so no event is skipped
  • Materialized view rebuilds run behind the serving alias; readers see no downtime

Rollback

  • Auto rollback triggers on consumer lag breach within the 48-hour watch window
  • Rollback is a consumer-group switch, not a redeploy, so it completes in under two minutes
  • Any data written by green during the window is corrected by the replay path