kuilt-otel-otlp
Send kuilt's offline-first telemetry to a standard OpenTelemetry collector.
When the network is available, this module forwards the spans, logs, and metrics kuilt buffered on the device to any OTLP/HTTP endpoint — the same collector your existing dashboards already read. It only sends what the endpoint has not already received, so a reconnect after hours offline uploads the gap, not the whole history.
The wire encoding is your choice: OTLP/JSON (the default, mature on every platform) or the more compact OTLP/protobuf that many collectors expect by default. Both carry the same data and reconcile the same way — set wire = OtlpWireFormat.PROTOBUF on OtlpHttpEdge to switch, and the requests go out as application/x-protobuf.
OtlpHttpEdge is the OtlpEdge a WarpOtlpBridge drains into. Point it at a collector URL and call drain on each reconnect.
Upgrading past this release: one burst of duplicates, once
Worth knowing before you see it in a dashboard.
"What have I already sent to this collector?" is a note kuilt keeps on the device, filed under a name built from the collector's address. This release rewrites that name twice over: entry names are now stored losslessly rather than squashed onto a narrower alphabet (#2506, #2511), and the part identifying the collector is now a fingerprint wide enough that two collectors cannot be mistaken for one (#2513). The old note is left where it lies rather than moved, so the first run after the upgrade finds no record of what was sent and offers up everything still sitting in the buffer.
What that looks like at the collector is one burst of records it has seen before. It is bounded by whatever the device's buffer still holds, it happens on the first flush only, and every flush after it is back to normal. So if telemetry doubles up right after an upgrade, this is that — not a retry loop worth chasing. Nothing is lost: the re-sent records carry the same ids they carried the first time, so a collector that deduplicates by id absorbs them, and the cost is bandwidth rather than double-counted data.
Both renames deliberately ship in the same release. Landing them separately would have handed consumers this burst twice instead of once.