Every observability tool ships as a database with charts bolted on. You write a query, wait, read a chart, write another query. During an incident, at 3 a.m., under pressure — this is the wrong interface.
Engineers don't reason about systems as queries. They reason about them temporally — what happened, what came before, what was doing what at the moment it broke. And what broke is rarely only in the telemetry. A commit landed an hour ago. A queue drained. A rule in a security group flipped. Every cause leaves a mark somewhere. The tool should gather all of them onto one axis.
Spoonfish does. A playhead. Tracks. Waveforms. Spans. A topology that behaves like a map. An axis wide enough for the whole system, not just its telemetry. No query bar. No tab switching. No correlation-by-screenshot.
A service graph drawn by a force-directed algorithm rearranges itself every time you open it. That's not a map — it's weather. Spoonfish renders topology the way cartographers do: stable geography, semantic zoom, a shape you can learn.
Zoom out for the shape of the estate. Zoom in until you can read the service names and watch a single request thread its way across them — one continuous route, picking up color as health or latency change along the hops. Pins where it mattered. The map doesn't reshuffle. It resolves.
Click any service. A dialog drops over the map with its logs for the same time window — anomaly-ranked, templates clustered, heartbeat faded. The line that matters comes with the neighbors that matter: the deploy that preceded it, the topology change that coincided with it, the sibling service whose error rate spiked alongside. Logs as a node in an event graph, not a stream in a silo.
- Transport
- Play, pause, scrub, zoom. Sub-millisecond seek across hours or weeks. The playhead is the only cursor that matters; every track below it moves in lockstep.
- Tracks
- Every moving part as a dedicated track — utilization waveforms, span timelines, error pins, deploy markers, commit landings, topology deltas, kernel events. Whatever your system can emit goes on the axis, next to everything else that moved with it.
- Cartography
- A topology built on the logic of a geographical map, not a force-directed graph. Services sit where you left them. Zoom out for the shape of the whole system; zoom in until individual routes, pins, and boundaries resolve. A map you can learn, not a diagram that reshuffles.
- Mute, Solo, Drill
- Hide the healthy. Isolate the suspect. Expand a track into its children. Debug by elimination, not by query — the same way you'd debug with headphones on.
- Precognition
-
The recording doesn't stop at
now. The same model that replays the last hour extrapolates the next few seconds, so you see the spike forming before the alert fires — a flight recorder that also looks ahead. - Full fidelity, your bucket
-
No sampling. No cardinality tax. High-cardinality dimensions are a feature, not a
billing line. Columnar
Parquetover yourS3orR2bucket, queried in-process withDataFusion. Open format, always exportable, never held hostage.
First cohort, Q3 2026.
Private beta opens with a small group of design partners — teams who've given up on correlating by screenshot and want the whole system on one axis. All stacks welcome; mixed stacks preferred.
One email when access opens. Nothing else.