Xray
daily rollups 1h 6h 24h 48h 3d 5d 7d 14d 30d ‹
Last 30 days▾
Ends now. Anything up to this cluster's 30-day retention.
Start and end times
Cancel
Interpreted in UTC (change it at the right of this bar). Earliest available: 2026-10-19 13:33:12.
› →|
updated 1d ago
Sample data. This database was generated by scripts/seed_dev_db.py — every service, deploy, person and figure on this page is synthetic. Names are sample.*@seed.invalid and image tags are prefixed SEED-. Nothing here happened.

What's where

X-ray reads your fleet's real logs and metrics — nothing here is sampled or synthetic. Every page is scoped by the cluster and time window in the bar above, so if a page looks empty, check those two first. Below: every page and the question it answers, then the three routes people actually take.

Start with the job, not the page

An endpoint got slow

  1. Open Traffic → API and sort by latency to find which endpoint, and confirm it is actually slower rather than just busier.
  2. Click through to Traffic → Traces for one slow request and read where the time went hop by hop.
  3. If the time is in the database, Neo4j → Slow queries shows the statements, already attributed back to that endpoint.
  4. Neo4j → Query trends answers the follow-up: is this shape newly slow, or newly popular? Those need different fixes.

Someone reports an error

  1. Traffic → Errors groups failures by cause, so you see whether it is one user or everyone.
  2. If it is a login problem, Traffic → Auth0 has the actual auth events — lockouts and failures included.
  3. Traffic → Logs for the raw lines; filter by their email or a trace id.
  4. If the whole service is unhappy rather than one request, Fleet → Reliability shows restarts and OOM kills.

The nightly tests had a bad night

  1. Tests — the run list is grouped by date; the pass-rate chart plots only runs that produced results.
  2. Open the run. If it is tagged a collapse, the environment fell over and the failures are one problem, not forty.
  3. Otherwise "broke in this run" names the suites that passed in the previous healthy run — those are the regressions worth chasing.
  4. Click a suite to see whether it has always been red or broke on a date. Those are different conversations.

The bill went up

  1. Costs → Overview first: it names the month's biggest movers against the same days last month, so you start from what actually changed rather than what is merely expensive.
  2. Costs → Search is the fastest way to ask about one thing by name — type autopilot, neo4j, gemini, a service, a cluster, or paste a bucket link to price that bucket alone.
  3. Anomalies for a day that broke its own pattern, and Commitments to check you are using the discounts already being paid for.
  4. If the answer is "the model spend", Costs → AI splits it by service, org and job.

What did one document cost to process?

  1. AI → Documents lists every document the fleet ran a model over, what it cost, and how far it got — uploaded, TMF pipeline, validated in a work order, indexed for search.
  2. Open one for its totals — tokens in and out, model time against wall time, and every call on a time axis from the run's start, so a slow step is visible rather than inferred.
  3. The same page names the work that cannot be attributed: operations that expect a document and named none. That list is the honest edge of this data, not a gap to ignore.

Is production what we think it is?

  1. ArgoCD → In sync? answers it in one sentence per environment. If everything matches git, you are done — that is the whole point of the page.
  2. If something drifted, the same page names it: which services are running an image a person put there, then ConfigMaps, Secrets and ExternalSecrets, then permissions and CRDs.
  3. Clusters shows what ArgoCD manages across every environment; Who changed what puts a name against each change.
  4. To read a change against its effect, Change timeline puts traffic, errors and latency either side of each deploy.

A screen feels slow, and nobody can say why

  1. Traffic → Journey starts from the person, not the endpoint: one browser session, screen by screen, with the Cypher each screen caused.
  2. Waste is the weekly version of the same question — which screen × call pairs fetch the same thing more than once, ranked by the Cypher that wasted.
  3. Its Example links open the exact session behind a row. They open the window that ledger sampled, which is not the range in the bar above.

I want to be told when this happens again

  1. Alerts — describe the condition in words; X-ray writes the query and shows you what it would have fired on before you save it.
  2. Every message is written for a reader rather than assembled from field names, so a card says what happened and to whom.
  3. System alerts is the other half: things X-ray notices on its own, including a source that has gone silent — absence is a finding, not a blank.

Every page

Perimeter

The edge before the fleet: who reached our Cloudflare front door and what the firewall stopped — WAF and DDoS blocks, where they came from, and the rules holding the line.

Costs

What the fleet costs, every penny tied to a project, cluster, namespace or workload — and what is heading where.

Overview/costs This month so far, where it is heading, what moved most, and anything the ledger flagged — the one screen to open before Billing. Search/costs/search One box over every dimension of the bill: a word (autopilot, neo4j, gemini), a service, a cluster, a workload — or paste a bucket link and see what that bucket costs. Explore/costs/explore Break cost down by anything — project, service, SKU, cluster, namespace, workload, resource — and drill by clicking; every row compared with the previous period. Commitments/costs/commitments The prepaid commitments: how much of each fee funded usage, what it funded, and the dollars that funded nothing. AI/costs/ai The AI line on the bill — Gemini, Vision, Translate — by model and token kind, reconciled with what X-ray's own AI log says each org used. Cursor/costs/cursor Cursor AI, Cloud Agents, and CI/CD workflow execution costs: active vs skipped pipelines, Bugbot reviews, and model usage. OpenAI/costs/openai OpenAI organization and project spending, API key usage, token volumes, and model costs. Anomalies/costs/anomalies Days a project × service × cluster cell sat far outside its own normal, each decomposed into what moved and lined up with what X-ray saw that day. Forecast/costs/forecast Month-end and 90-day projections per project and cluster, against the monthly targets you set here. Recommendations/costs/recommendations Ways to spend less, labelled by who says so: Google's Recommender, X-ray's own rules over the ledger, or a model's analysis on request.

Admin

Fleet-owner tooling. Useful to whoever runs the platform, noise to everyone else — which is why it is out of the main bar.

Two things worth knowing

Every view shows the queries behind it. The button at the bottom of each page lists the exact BigQuery, Monitoring and GCS reads that produced what you are looking at — so a number you doubt can be checked against the source rather than taken on trust.

Absence is not zero. Where data is missing X-ray says so rather than plotting a zero: a build that produced no tests, a service with no trace ids, a cluster with no Neo4j log. A zero would read as "measured, and it was nothing", which is a different and much more alarming claim.