controlling-costs

SKILL.md

Axiom Cost Control

Dashboards, monitors, and waste identification for Axiom usage optimization.

Before You Start

Load required skills:

skill: axiom-sre
skill: building-dashboards

Building-dashboards provides: dashboard-list, dashboard-get, dashboard-create, dashboard-update, dashboard-delete

Find the audit dataset. Try axiom-audit first:

['axiom-audit']
| where _time > ago(1h)
| summarize count() by action
| where action in ('usageCalculated', 'runAPLQueryCost')

If not found → ask user. Common names: axiom-audit-logs-view, audit-logs

If found but no usageCalculated events → wrong dataset, ask user

Verify axiom-history access (required for Phase 4):

['axiom-history'] | where _time > ago(1h) | take 1

If not found, Phase 4 optimization will not work.

Confirm with user:

Deployment name?

Audit dataset name?

Contract limit in TB/day? (required for Phase 3 monitors)

Replace <deployment> and <audit-dataset> in all commands below.

Tips:

Run any script with -h for full usage

Do NOT pipe script output to head or tail — causes SIGPIPE errors

Requires jq for JSON parsing

Use axiom-sre's axiom-query for ad-hoc APL, not direct CLI

Which Phases to Run

User request
Run these phases

"reduce costs" / "find waste"
0 → 1 → 4

"set up cost control"
0 → 1 → 2 → 3

"deploy dashboard"
0 → 2

"create monitors"
0 → 3

"check for drift"
0 only

Phase 0: Check Existing Setup

# Existing dashboard?
dashboard-list <deployment> | grep -i cost

# Existing monitors?
axiom-api <deployment> GET "/v2/monitors" | jq -r '.[] | select(.name | startswith("Cost Control:")) | "\(.id)\t\(.name)"'

If found, fetch with dashboard-get and compare to templates/dashboard.json for drift.

Phase 1: Discovery

scripts/baseline-stats -d <deployment> -a <audit-dataset>

Captures daily ingest stats and produces the Analysis Queue (needed for Phase 4).

Phase 2: Dashboard

scripts/deploy-dashboard -d <deployment> -a <audit-dataset>

Creates dashboard with: ingest trends, burn rate, projections, waste candidates, top users. See reference/dashboard-panels.md for details.

Phase 3: Monitors

Contract is required. You must have the contract limit from preflight step 4.

Step 1: List available notifiers

scripts/list-notifiers -d <deployment>

Present the list to the user and ask which notifier they want for cost alerts.
If they don't want notifications, proceed without -n.

Step 2: Create monitors

scripts/create-monitors -d <deployment> -a <audit-dataset> -c <contract_tb> [-n <notifier_id>]

Creates 3 monitors:

Total Ingest Guard — alerts when daily ingest >1.2x contract OR 7-day avg grows >15% vs baseline

Per-Dataset Spike — robust z-score detection, alerts per dataset with attribution

Query Cost Spike — hardened z-score with 30d baseline, 5d exclusion gap, persistence-based gating (median_z > 3, p25_z > 2.5)

The spike monitors use notifyByGroup: true so each dataset triggers a separate alert.

See reference/monitor-strategy.md for threshold derivation.

Phase 4: Optimization

Get the Analysis Queue

Run scripts/baseline-stats if not already done. It outputs a prioritized list:

Priority
Meaning

P0⛔
Top 3 by ingest OR >10% of total — MANDATORY

P1
Never queried — strong drop candidate

P2
Rarely queried (Work/GB < 100) — likely waste

Work/GB = query cost (GB·ms) / ingest (GB). Lower = less value from data.

Analyze datasets in order

Work top-to-bottom. For each dataset:

Step 1: Column analysis

scripts/analyze-query-coverage -d <deployment> -D <dataset> -a <audit-dataset>

If 0 queries → recommend DROP, move to next.

Step 2: Field value analysis

Pick a field from suggested list (usually app, service, or kubernetes.labels.app):

scripts/analyze-query-coverage -d <deployment> -D <dataset> -a <audit-dataset> -f <field>

Note values with high volume but never queried (⚠️ markers).

Step 3: Handle empty values

If (empty) has >5% volume, you MUST drill down with alternative field (e.g., kubernetes.namespace_name).

Step 4: Record recommendation

For each dataset, note: name, ingest volume, Work/GB, top unqueried values, action (DROP/SAMPLE/KEEP), estimated savings.

Done when

All P0⛔ and P1 datasets analyzed. Then compile report using reference/analysis-report-template.md.

Cleanup

# Delete monitors
axiom-api <deployment> GET "/v2/monitors" | jq -r '.[] | select(.name | startswith("Cost Control:")) | "\(.id)\t\(.name)"'
axiom-api <deployment> DELETE "/v2/monitors/<id>"

# Delete dashboard
dashboard-list <deployment> | grep -i cost
dashboard-delete <deployment> <id>

Note: Running create-monitors twice creates duplicates. Delete existing monitors first if re-deploying.

Reference

Audit Dataset Fields

Field
Description

action
usageCalculated or runAPLQueryCost

properties.hourly_ingest_bytes
Hourly ingest in bytes

properties.hourly_billable_query_gbms
Hourly query cost

properties.dataset
Dataset name

resource.id
Org ID

actor.email
User email

Common Fields for Value Analysis

Dataset type
Primary field
Alternatives

Kubernetes logs
kubernetes.labels.app
kubernetes.namespace_name, kubernetes.container_name

Application logs
app or service
level, logger, component

Infrastructure
host
region, instance

Traces
service.name
span.kind, http.route

Units & Conversions

Scripts use TB/day

Dashboard filter uses GB/month

Contract
TB/day
GB/month

5 PB/month
167
5,000,000

10 PB/month
333
10,000,000

15 PB/month
500
15,000,000

Optimization Actions

Signal
Action

Work/GB = 0
Drop or stop ingesting

High-volume unqueried values
Sample or reduce log level

Empty values from system namespaces
Filter at ingest or accept

WoW spike
Check recent deploys