Skip to content
ACloud & DevOpsService 04 · B4

Deploys that areuneventful.

Reproducible environments, a release path that runs without ceremony, observability from the first deploy, and a cost profile you can forecast. Infrastructure treated as a product with an owner, rather than a pile of consoles.

X
ReproducibilityEnvironments defined in code, rebuilt on demand.
Y
VisibilityKnowing what is happening before a customer tells you.
BEnvironmentsOne definition, three lanes
Release pathB — figure 01
ENVIRONMENTSDevelopmentREBUILT ON DEMANDStagingIDENTICAL BY CONSTRUCTIONProductionPROGRESSIVE ROLLOUTONE DEFINITION IN CODETESTSREVIEWGATEDEPLOYEDLOGS · METRICS · TRACES
Environment laneGate and productionObservability strip
Every lane is built from the same definition in code, which is what makes testing in staging mean anything. The release steps through gates rather than around them, and the strip underneath is how you know it worked.

A deploy should be the least interesting part of a Thursday.

That is not achieved by being careful. It is achieved by removing the steps that require care: the manual configuration, the environment only one machine can reproduce, the check that lives in a single person’s memory.

  1. 01

    Environments

    identical by construction

  2. 02

    Deployment

    one path, automated

  3. 03

    Infrastructure

    declared, reviewed

  4. 04

    Observability

    logs, metrics, traces

  5. 05

    Scaling

    planned, costed

CSymptomsWhat a bad release looks like
01

Releases are events

Scheduled for a Friday evening, performed by one person, followed by a nervous weekend. The goal is a release too boring to schedule.

02

Environments have drifted

Staging no longer resembles production, so testing there proves very little. Environments defined in code stop the drift at its source.

03

You find out from a customer

Either no alerting, or so much alerting that nobody reads it. Instrumentation belongs in the first deploy, with a small number of signals that mean something.

04

The bill is a surprise

Cloud spend without an owner drifts upward quietly. A forecast, a budget alert and a periodic review make it a decision instead of a discovery.

DWhat we build4 shapes of engagement
D1

Environments as code

Every environment described in a repository and rebuildable from scratch, including the parts that were originally clicked into existence.

  • Infrastructure as code
  • Secrets handling
  • Ephemeral environments
  • Reviewed changes
D2

A release path

Build, test, deploy, verify — one automated path per service, with a rollback that has actually been used.

  • CI pipelines
  • Automated gates
  • Progressive rollout
  • Tested rollback
D3

Observability

Structured logs, the handful of metrics that describe health, traces where a request crosses services, and alerts a human can act on.

  • Structured logging
  • Health metrics
  • Tracing
  • Actionable alerts
D4

A scaling and cost plan

What breaks first, at what load, and what it will cost — written down before the growth arrives.

  • Load profile
  • Bottleneck analysis
  • Budget alerts
  • Capacity plan
EObservability4 signals

Four signals, and a reason to wake someone for each.

Alerting that nobody reads is the same as no alerting. We instrument a small, deliberate set of signals, agree the threshold that justifies an interruption, and route everything else to a dashboard nobody has to watch.

What we watch4 signals
  • Availability

    Whether requests are being served

    Alert — Sustained error rate

  • Latency

    Whether they are served in time

    Alert — p95 beyond the budget

  • Saturation

    How much headroom is left

    Alert — Approaching a known limit

  • Freshness

    Whether the data behind it is current

    Alert — A pipeline missed its window

FApproachHow the work runs
  1. 01

    Make one environment reproducible

    Usually the least glamorous one. Everything after this is easier once a rebuild is a command rather than an archaeology project.

  2. 02

    Automate the path to production

    The same steps every time, in the same order, with the checks that used to live in someone's memory encoded as gates.

  3. 03

    Instrument before you need it

    A small, deliberate set of signals, wired up while the system is calm rather than during an incident.

  4. 04

    Rehearse the failure

    Restore the backup. Roll back the release. Lose the node. A recovery plan that has never been run is a hypothesis.

L1Platforms
AzureAWSManaged databasesCDN and edge
L2Packaging
ContainersRegistriesBuild cachingArtefact versioning
L3Automation
Infrastructure as codeCI/CDEnvironment promotionSecrets rotation
L4Operations
LoggingMetricsTracingBackup and restore drills
GContact06 Clarity · resolved

Tell us what your last release felt like.

If it involved a checklist, a phone call and a held breath, there is a shorter path. We will map the one you have, and the one you could have.