The classifier DLQ (apm-wo-analysis-classifier-dlq) had no alarm, so a
message landing there after Lambda exhausts async retries — a genuinely
dropped classifier run — surfaced nowhere.
Add a CloudWatch alarm on AWS/SQS ApproximateNumberOfMessagesVisible
(Maximum, threshold >0, 300s period, 1 evaluation period, notBreaching)
that pages the shared site-alerts SNS topic. Mirrors the
workorder-email-processor-dlq-messages alarm in procurement-ingest's
wo_stack and the payments-payroll-batch-dlq-messages alarm convention.
ALARM-only, no OK action, per the CloudWatch-alarm preference.
Org pin moved from 2.253.1 (bundles vulnerable fast-uri 3.1.0, 2 high
GHSAs) to 2.257.0 (bundles patched 3.1.2). Removes the blanket
dependabot ignore per the new handbook pinning policy (exact pins kept
current by Dependabot; blanket ignores banned).
From the cross_reviewer (GPT-4.1) per-PR passes, now that the orchestrator is
back up:
Phase 2 (classifier):
- Process ALL S3 records, not just event["Records"][0] — batched notifications
no longer silently dropped (the review's only BLOCK).
- Derive the partition dt from the S3 event time, not the Lambda wall-clock —
stable across retries / the midnight boundary.
- Add an SQS dead-letter queue so a failed run surfaces instead of dropping a
day's data after Lambda's retries.
Phase 4 (Slack):
- Stage throttling (rate 10 / burst 20) on the public /slack/interactions HTTP
API. (AWS WAF doesn't attach to apigwv2 HTTP APIs; stage throttling is the
mechanism.)
Phase 5 (Grafana):
- Explicit encrypted=True on the gp3 root volume.
Tests: synth assertions for the DLQ, stage throttling, and the encrypted volume.
60/60 pass; cdk synth green for both stacks. Deferred NITs (print->logging, sig-
failure source-IP logging, S3 versioning, CIDR-maintenance runbook) -> Phase 6.
NOTE: like the earlier deploy fixes these sit on phase-5 but span phases — the
classifier/DLQ to #8, throttling to #10, encryption to #11 — reconcile at merge.
The encrypted-volume change needs the deferred clean instance replacement to
take effect (can't encrypt a live volume in place).
aws s3 sync skips same-size files on download unless --exact-timestamps is set,
so a dashboard edit that doesn't change file size (e.g. refresh 2->1, or a query
tweak) never propagated to the instance. Add --exact-timestamps to all four
sync invocations (boot + 15-min timer).
Two issues found loading the deployed dashboard:
1. Panels used type "bar-chart" (hyphenated); Grafana's core panel is "barchart"
— hence "plugin bar-chart required". Fixed both panels.
2. ALL panels showed "no data" because Athena failed with HIVE_BAD_DATA:
the classifier wrote summary.json/details.json INTO analytics/dt=*/ — the
same prefix the Glue table scans — so Athena tried to read the JSON as
Parquet and every query failed. Move the metadata to a separate meta/dt=*/
prefix: classifier writes there (grant_read_write meta/*), the Slack Lambdas
read there (read_meta_json, grant_read meta/*), and analytics/ holds only
Parquet. Verified: the category GROUP BY query now succeeds against Athena.
Found on first boot (cloud-init errored, grafana-server never started):
- grafana-cli needs --homepath=/usr/share/grafana or it can't find config
defaults; under set -e that aborted the whole bootstrap.
- pinned plugin version 2.18.2 doesn't exist (conflated with awswrangler's
version) — grafana-athena-datasource latest is 3.2.0.
Verified by running the corrected bootstrap on the instance via SSM: plugin
installs, grafana-server active, /api/health 200, ALB target healthy.
NOTE: the running instance was repaired in-place (the user-data change updated
the launch template but did not replace the instance). The committed user-data
is now correct, so a fresh launch boots clean — a one-time clean instance
replacement should validate that before prod sign-off.
Two issues only a real deploy/run surfaced (synth + offline tests passed):
1. Classifier exceeded Lambda's 250 MB unzipped limit (bundled awswrangler +
pandas + pyarrow + numpy). Move them to the AWS-managed SDK-for-pandas layer
(AWSSDKPandas-Python312-Arm64:27, awswrangler 3.16.1, pre-stripped to fit);
bundle only openpyxl. Drop the unused anthropic SDK — _call_haiku uses stdlib
urllib. Function package now ~890 KB.
2. Slack rejected the daily post with invalid_blocks: every category drill
button shared action_id "drill_category". Qualify it as "drill_category:<cat>"
for uniqueness; the interactions handler now matches on the prefix. Add a
regression test asserting all daily-summary action_ids are unique.
Verified in prod: classifier writes Parquet + summary.json + details.json;
slack-post posts the daily summary + 3rd-escalation alert; the interactions
endpoint (apm-wo.seahaven.com) returns 401 on a bad signature. 58/58 tests pass.
NOTE: these fixes sit on the phase-5 branch but logically belong to earlier
phases — the layer fix to #8 (classifier), the Slack fix to #10 — and must be
moved/cherry-picked there before those PRs merge independently. See cleanup.
The one non-serverless piece — Grafana OSS on a t4g.small (AL2023, ARM64) in the
imported seahaven-vpc, fronted by an internet-facing ALB locked by SG to the
office CIDRs (no Client VPN exists, so "VPN-only" = office-IP restriction, the
syslog-server pattern). Instance in private subnets, reachable only from the ALB
SG, administered via SSM Session Manager (no SSH/key pair).
grafana_stack.py: ALB (HTTPS, *.seahaven.com cert, open=False so the SG office
rules aren't undone by an auto 0.0.0.0/0), instance role (Athena query + Glue
read + S3 analytics/athena-results, no static keys), Route53 grafana.seahaven.com
alias, gp3 root volume RETAINed, daily DLM snapshot of the tagged instance, and a
BucketDeployment that uploads grafana/ to the S3 config prefix.
grafana_userdata.sh: install Grafana OSS, pin the Athena datasource plugin, write
grafana.ini (root_url grafana.seahaven.com, kiosk embedding), sync provisioning +
dashboards from S3 on boot, and a systemd timer re-syncs every 15 min so repo
edits land without an instance rebuild.
Dashboard (grafana-author agent, grafana/dashboards/apm-work-orders.json, uid
apm-wo so the Slack 📊 button resolves): 7 panels — category distribution,
escalation summary, action/routine, escalations-by-site, trend time-series over
dt (the new capability), filterable WO table (5 template vars, escalation row
coloring, CSV export, no APM links), and the mismatch panel. Datasource uid
"athena" pinned in the provisioning yaml.
Tests: tests/test_grafana_synth.py — ALB admits only the office CIDRs on 443
(caught and fixed a default 0.0.0.0/0 listener rule), instance only-from-ALB,
no static keys, scoped instance role + SSM, gp3+retained root volume, daily DLM
backup, grafana.seahaven.com alias. 57/57 tests pass; full cdk synth green.
Two push surfaces (no App Home) + interactive drill-down, per CLAUDE.md.
Block Kit (blockkit.py, pure/offline): build_daily_summary (header, vs-yesterday
deltas, escalation breakdown with 3rd highlighted, action/routine, top sites,
mismatch callout, category drill buttons + 📊 Open dashboard link, footer),
build_escalation_alert (one @here, returns None on zero-3rd — suppression), and
build_wo_modal (views.open payload, capped under Slack's 100-block limit).
Lambdas: slack_post/handler.py (classifier-invoked: read today/yesterday
summary.json, post daily summary, conditionally post the batched alert from
details.json) and slack_post/interactions.py (API Gateway: verify Slack
signature, filter details.json, views.open the WO modal within the 3s trigger_id
window). slackio.py centralizes Secrets Manager creds, the SSM dashboard URL,
signature verification, and analytics/ reads — keeping blockkit pure.
Classifier: emit analytics/dt=*/details.json (per-WO index for the modals) and
async-invoke slack-post after the snapshot write (best-effort; a Slack failure
never fails classification).
CDK: slack-post + interactions Lambdas (Docker-bundled slack_sdk), HTTP API on
apm-wo.seahaven.com (wildcard ACM cert + Route53 alias; signature-verified, so
the route is unauthenticated by design), SSM /apm-wo-analysis/grafana-base-url,
and scoped IAM (read analytics/, read the Slack secret + dashboard param;
classifier granted lambda:InvokeFunction on slack-post). Slack creds live in one
Secrets Manager secret apm-wo-analysis/slack-credentials {botToken, signingSecret,
channelId}; cdk.json gains cert/zone/domain context.
WO drill-downs link to Grafana only — no APM deep-links (per decision).
Deliverables for test time: slack/manifest.yaml (app manifest, interactivity
request_url = apm-wo.seahaven.com).
Tests: tests/test_blockkit.py (30 offline cases — deltas, zero-3rd None, <100
blocks under large inputs, modal truncation/overflow, dashboard URL) and Phase 4
assertions in test_pipeline_synth.py (both Lambdas, the API route/domain/alias,
and no broad/write IAM on the Slack roles). 49/49 tests pass; cdk synth green.
CDK-define the apm_wo_snapshots Glue table with partition projection over
analytics/ (dt as projected date partition, 2026-01-01..NOW). Projection means
no crawler, no MSCK REPAIR, and — critically — the classifier needs no Glue
catalog access at all.
pipeline_stack.py: glue.CfnTable (Parquet SerDe, 17-column schema mirroring the
classifier's snapshot incl. contractor_description and the two boolean flags) +
athena.CfnWorkGroup `apm-wo-analysis` (enforced result location, SSE-S3) + a
30-day lifecycle rule on athena-results/ (disposable query output in a RETAIN
bucket). Trim the classifier role: drop the entire Glue policy statement.
handler.py: stop registering the table at runtime — drop database=/table= from
to_parquet so the classifier writes pure Parquet; partition projection handles
the rest. Keeps overwrite_partitions for idempotent same-day re-uploads.
test_pipeline_synth.py: offline synth assertions (bundling skipped) — projection
properties, column schema/types, workgroup result enforcement, and that no IAM
policy grants glue:* to the classifier.
cdk synth green; 11/11 tests pass.
Implement the core classification engine and wire it into the pipeline stack.
classify.py: two-axis classifier — HTML-strip, comment-intent regex buckets
(escalations → status inquiry, most-specific first), Hold Reason / WO Status
structured state, comment-vs-state mismatch detector, and a Claude Haiku
fallback (Secrets Manager key) reserved for ambiguous free-text. Exports
ESCALATION_CATEGORIES / ACTION_NEEDED_CATEGORIES.
handler.py: S3-triggered handler — parse xlsx/csv, classify each non-blank
row, write a per-WO Parquet snapshot to analytics/dt=YYYY-MM-DD/ (registers
the Glue partition via awswrangler) and a summary.json for slack-post (Phase 4).
pipeline_stack.py: Glue database, ARM64 Python 3.12 classifier Lambda
(Docker-bundled deps), S3 raw/ notification (.xlsx/.csv), and least-privilege
IAM (read raw/, read-write analytics/, scoped Glue catalog, read Anthropic key).
Smoke-tested against the real export: 347 rows, "Other" at 5.2% (target ~9%),
18 mismatches flagged. 7/7 unit + smoke tests pass; cdk synth green.
Complete Phase 1 ingestion. Add a least-privilege IAM user
(apm-wo-drop-uploader) to the pipeline stack, scoped to s3:PutObject
on the raw/ prefix only — the local launchd uploader authenticates as
this user via a dedicated profile, so a laptop credential leak cannot
read, list, or touch the analytics data.
Replace the scaffold uploader stub with the hardened stampli-pattern
script (lockfile, logging, timestamped archive, notifications, settle
delay) and align names to the convention (~/apm-wo-drop, ~/.local/bin,
com.seahaven.apm-wo-uploader). The plist sets PATH/HOME because launchd
runs with a stripped environment and otherwise cannot find aws.
The exports bucket already shipped in the Phase 0 scaffold, so the code
delta here is the uploader identity and tooling.
Stand up the Phase 0 CDK scaffold for the daily APM work-order
analysis pipeline: two-stack CDK app (pipeline + grafana), classifier
and slack-post Lambda packages, dashboards-as-code, the local
drop-folder uploader, and a classifier smoke-test placeholder.
Wire CI/CD to the org reusable workflows: ci.yaml -> ci-python-sam
(ruff + cdk synth) and deploy.yaml -> cd-cdk (OIDC, cdk deploy --all).
Pin aws-cdk-lib==2.253.1; Lambdas target Python 3.12 / arm64.
Rewrite .gitignore to the org Python-CDK standard so the source-of-
truth files (CLAUDE.md, docs/, .claude/agents) are tracked while build
artifacts (.venv, cdk.out, caches) stay ignored.
Domain logic, stack resources, and dashboards are stubbed and filled
in across Phases 1-5 (docs/BUILD.md). cdk synth is green for both
stacks; ruff check/format pass.
