Advisor

What is slow or wasteful in your setup, from your own measurements: CI caches, rebuilds, idle machines, missing indexes, with the fix.

Advisor shows what is slow or wasteful in your own setup on Deutsche Cloud and how to fix it: a CI workflow that downloads every dependency on every run, a redeploy that builds an image it already has, a machine that has done nothing for two weeks, a query that scans a table on every call. Each finding says what was measured, what it costs you in minutes or euros a month, and the fix. Where the fix is a setting Deutsche Cloud owns, Apply makes it; where it is a file in your repository, Open a pull request proposes it; anything else links to the resource or explains the steps. Console: Management → Advisor (/advisor).

Checking your own setup is included; nothing in Advisor is billed.

Where findings come from

  • Rules, every hour and on request (Check now, dcloud advisor evaluations start). A rule is a fixed check over the platform's own measurements: run and step timings, deployment phases, CPU and network per minute, query statistics, what billing charged. Rules are exact and always on.
  • AI review, once a day, for organizations that allow AI to read live data (Account → AI settings → Read live data). A model reads the measurements of your busiest CI repositories (at most ten a day) and their workflow files, and proposes findings no rule covers: the same build done twice, jobs that could run in parallel, a matrix bigger than it needs. What it sees: the measured numbers and the workflow files with every literal env: and with: value removed. Never logs, never secret values. A proposal is kept only when every number it cites is one of those measurements and its claimed saving is no larger than the durations it cites; its confidence is the model's own statement and is shown as such. Review now on the page, or dcloud advisor evaluations start-ai-review, runs it at once (at most every ten minutes). Each review is in your audit log.

What it checks

Check Service What it looks at Fix
Dependency cache in CI CI An npm/pnpm/yarn/pip/poetry/go/cargo install with no cache, taking ≥ 20 s (median, ≥ 3 runs in 30 days) Pull request
Cache keys that never match CI Workflows whose cache lookups missed on ≥ 80 % of their last runs (≥ 5 runs) Steps
Configuration changes without a rebuild Apps Redeploys of the commit already running that built and pushed the image again (≥ 60 s) Steps
Health-check timeout that fits the app Apps A timeout set to ≥ 180 s and ≥ 5 × the app's 95th-percentile start time (≥ 5 rollouts) Apply
Idle machines Compute CPU averaged under 2 % with no 5-minute sample above 10 % for 7+ days Apply (stop)
Machines bigger than their load Compute A smaller size of the same family fits the 95th-percentile CPU with headroom On the resource
Long-stopped machines whose disks still bill Compute Stopped 7+ days with volumes attached On the resource
Volumes attached to nothing Compute Detached 7+ days On the resource
Elastic IPs nobody uses Compute Associated with nothing for 24+ h On the resource
Databases bigger than their load Databases 95th-percentile CPU under 10 % per vCPU for 7+ days On the resource
Slow queries with a missing index Databases Statements reading ≥ 100 pages per row returned and taking ≥ 5 ms (Query Insights, 7 days) On the resource
Connections close to the limit Databases ≥ 10 minutes in 7 days with client connections at ≥ 80 % of max_connections Steps
Alarms that notify someone Monitoring Alarms you made with no e-mail address and no webhook On the resource
A budget on the bill Billing No budget at all while the last 30 days were charged €20 or more On the resource
No single issue drowning the rest Error Tracking One unresolved issue with ≥ 50 % of a project's events in 7 days (≥ 1 000 events), from at most 10 users On the resource
Errors refused by a limit Error Tracking ≥ 1 % (and ≥ 100) of the events sent in 7 days answered 429 by a project, key or organization limit On the resource
A rate limit on bursty projects Error Tracking No per-minute limit on the project or any key, ≥ 10 000 events in 7 days, busiest hour ≥ 10 × the usual hour Apply
Databases nobody uses Databases On every day of the last 7–14, at most our own probe connected (daily 95th percentile ≤ 1) and CPU averaged under 5 % On the resource
Database storage that will fill up Databases The disk's growth over the last 72 hours reaches the size it may grow to in under 14 days, or it is ≥ 80 % full with no autoscaling room On the resource
Machines, databases and apps with an alarm Monitoring A running machine, database or app, a day old or more, with no enabled alarm on it Apply
Uptime checks that notify someone Monitoring Uptime checks with no e-mail address and no webhook On the resource
Scheduled workflows that keep failing CI The last ≥ 5 scheduled runs of a workflow all failed On the resource
Frequent schedules on quiet repositories CI A schedule that started more than 24 runs a day for the last 7 days on a repository with no push for 7 days Steps
Job timeouts where runs have hung CI A job without timeout-minutes in a workflow whose run timed out or ran ≥ 30 min and ≥ 5 × its usual length Steps
Runs waiting for a runner CI Queued-to-started wait with a 90th percentile over 2 minutes (≥ 10 runs in 7 days) Steps
Self-hosted runners that are really gone CI A runner shown online that has not reported for an hour or more On the resource
Cache keys that change on every run CI Each of a workflow's last 5 runs asked for a cache key no other run asked for, and none was found Steps
Slow release commands Apps The release command's median over ≥ 3 rollouts in 30 days is a minute or more Steps
Environment changes waiting for a deploy Apps A variable, secret or binding changed over an hour after the running deployment started, with no rollout since On the resource
Failing messages with a dead-letter queue Queues A queue with no dead-letter queue holding messages received 5 times or more On the resource

Security findings live in Security Posture; the Advisor page shows its open findings by severity and links there, for anyone who may read the posture summary.

The cost checks are the same rules as the Cost explorer's savings: a machine is idle by the same definition in both places. Machines that belong to an app, a load balancer, a scaling group, a Kubernetes node pool or a CI runner pool are never called idle or oversized here; their owner has its own fix.

Resources too new or with too little data to judge are listed under Checks with the reason ("2 of 5 runs"), never silently left out.

The numbers

  • Minutes a month come from measured step and phase timings over the last 30 days. Where the saving is a projection (how much of an install a cache saves), the finding says estimate and how it was computed.
  • Euros a month are what billing actually charged or stops charging at your prices, never a list price. A stopped machine is not billed; its volumes still are, and the finding says so.
  • Minutes and euros are never added together. A fix that would raise your bill says so with its price.

Applying a fix

Every fix runs as you:

  1. Apply needs advisor:ApplyRecommendation on the finding and the resource's own action (for a stop, ec2:StopInstances on the machine). The check runs again first: if it no longer finds the problem, nothing changes and you get a 409. The change is made through the resource's own API, so it is in the audit log under your name exactly as a click on the resource's page. Advisor only applies changes you can undo; deleting a volume or releasing an address is done on the resource's page.
  2. Open a pull request needs git:PushRepository and git:CreatePullRequest on the repository. The change is committed to a new branch advisor/<check>-<id> and opened as a pull request against the default branch; your CI runs on it like on any pull request. The edit keeps your file as it is and only adds lines (for a cache: one actions/cache step before the install).
  3. Explain with the assistant opens the console assistant with a question about the finding; it reads the finding and can draft a change, which it labels as a draft.

Accepting and dismissing

Accept says "useful, we will fix it"; Dismiss takes a reason (on purpose, not useful, will not fix, wrong) and an optional note. A dismissed finding opens again when its numbers get twice as bad. A finding resolves by itself when a later check no longer finds it. Your answers rank later findings of the same check higher or lower in your organization; they never hide one. Dismissing a kind of AI review finding as not useful three times mutes that kind for your organization; the page lists muted kinds with an Unmute.

Who sees what

advisor:ListFindings and advisor:GetFinding are ordinary reads, but every finding, count and badge is shown only to someone who may read the resource it is about: a CI finding needs ci:ListRuns on the repository, a machine finding ec2:DescribeInstances, a query finding rds:DescribeDBQueryInsights. A finding you may not read answers 404, as if it did not exist.

Action What it allows
advisor:ListFindings List findings and the status of the checks
advisor:GetFinding Read one finding
advisor:ApplyRecommendation Apply a fix or open its pull request (plus the resource's own action)
advisor:UpdateFinding Accept, dismiss or reopen a finding
advisor:StartEvaluation Run the checks or the AI review now
advisor:ManageAiReview Mute or unmute a kind of AI review finding

API and CLI

dcloud advisor list                           # open and accepted findings, biggest saving first
dcloud advisor findings list --category cost
dcloud advisor findings get <id>
dcloud advisor findings apply <id>
dcloud advisor findings update <id> --state dismissed --reason intended --note "standby box"
dcloud advisor evaluations start

The same operations are under /advisor/api (GET /advisor/api/findings, POST /advisor/api/findings/{id}/apply, …) and available to AI assistants connected over MCP.

Did this page help you?