Skip to content

Field report: production dogfood observations (latency, version stamping, auth defaults) — 2026-08-14 #2

Description

@awksedgreep

Observational notes from two days of incident forensics against the production timeless-stack on ddnet dhcp1 (0.7.2, all three planes). Nothing here blocks; capturing per the file-everything policy.

Latency: unqualified pass

Every LogsQL query issued — dozens across two days — returned sub-second, including phrase scans over storm-hour data (6k+ matching rows), extract pipes, and count_uniq aggregations, all under live production write load (metrics + logs + traces ingesting simultaneously, oldest_queued_ms: 0 throughout). For a v0.7.x engine on real workload, response feel is already in the "instant" class that drives tool adoption. The known ceilings (work-row limit: timeless_logs#2, term intersection: timeless-libsql#2) are policy/planner issues, not raw-speed issues.

Storage efficiency at the same time: ~14.7 B/point (metrics, excluding WAL), ~29 B/log entry (~10x on repetitive DHCP logs), ~40 B/span vs ~232 B/span ingested.

Version stamping anomaly in the publish pipeline

On dhcp1's root podman storage: ghcr.io/awksedgreep/timeless-stack:latest (image ce7c1661, built ~13h before observation) self-reports 0.7.2 in /health — while 0.7.6-local (cf106bbb) was built an hour earlier. A :latest image built after 0.7.6 carrying 0.7.2 metadata suggests the publish pipeline stamps version from a stale source (old tag/env) rather than the built tree. Consequence: pull-and-restart on dhcp1 (2026-08-13) looked successful but still served 0.7.2. A timeless_build_info metric (timeless_stack#1) would make this visible on a dashboard; fixing the stamping makes it not happen.

Data-plane auth default caused silent ingest failure in dev

The dev instance's auth-on-by-default produced continuous HTTP 401 data_plane_authorization / missing_credentials from ddnet's log shipper — logs silently dropped rather than queued (client-side half tracked as ddnet#1). Known and being handled separately; recorded here because "default config silently loses data in the simplest local setup" is the kind of first-touch experience that costs early adopters.

Error taxonomy: positive note

{"error":"unsupported_capability","reason":"..."} with specific messages ("logical operator "or" is not implemented yet") is genuinely good DX — honest about limits, instantly actionable. The one inconsistency found (uppercase OR bypassing that honesty) is timeless_logs#3.

🤖 Generated with Claude Code

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions