# gRPC senders: use the upstream OpenTelemetry Collector

Bucker has **no gRPC door**. Its OTLP surface is OTLP/HTTP only (`/v1/traces`, `/v1/logs`,
`/v1/metrics`; see [otlp.md](otlp.md)). That is deliberate: a second listener and protocol stack
would carry its own authentication, limits, quota and outcome accounting, and every one of those
would have to agree with the HTTP door forever.

A sender that can only speak OTLP/gRPC (a stock `OTEL_EXPORTER_OTLP_PROTOCOL=grpc` SDK, a Go or
Java service that defaults to gRPC, a fleet agent) sends to the **upstream OpenTelemetry
Collector** instead, and the Collector forwards over OTLP/HTTP. Use the plain `otlp` receiver, the
`batch` processor and the `otlphttp` exporter. No contrib components, no Bucker-specific build.

## OTLP-only configuration

```yaml
receivers:
  otlp:
    protocols:
      grpc:
        endpoint: 0.0.0.0:4317
      http:
        endpoint: 0.0.0.0:4318

processors:
  batch:
    # Keep every request inside Bucker's published limits (otlp.md#limits): the door refuses a
    # request over 5,000 records or 4 MiB whole, rather than trimming it.
    send_batch_size: 2000
    send_batch_max_size: 4000
    timeout: 5s

exporters:
  otlphttp/bucker:
    endpoint: https://us.ingest.bucker.io          # your DSN's ingest host; no path
    headers:
      Authorization: "DSN ${env:BUCKER_DSN}"
    encoding: proto
    compression: gzip

service:
  pipelines:
    traces:
      receivers: [otlp]
      processors: [batch]
      exporters: [otlphttp/bucker]
    logs:
      receivers: [otlp]
      processors: [batch]
      exporters: [otlphttp/bucker]
    metrics:
      receivers: [otlp]
      processors: [batch]
      exporters: [otlphttp/bucker]
```

The `otlphttp` exporter appends `/v1/traces`, `/v1/logs` and `/v1/metrics` to `endpoint` itself.
The DSN is supplied from the Collector's environment and never written into the file.

Then point the gRPC sender at the Collector (`http://collector:4317`), not at Bucker.

## Limits the Collector must respect

| Limit | Value | Collector setting |
| --- | --- | --- |
| Records per request | 5,000 | `send_batch_max_size: 4000` (a safety margin) |
| Request body | 4 MiB | `compression: gzip` shrinks the wire size, but the 4 MiB limit also applies to the decompressed body, so keep batches modest |
| Attributes kept per collection | 256 | cut by the server; no setting |
| Shared (resource/scope) attribute bytes | 16 KiB | cut by the server; keep resource attributes small |

The Node SDK's own bridge (`initOtel({ logs: true, metrics: true })`) applies the same limits itself
before sending, splitting or truncating a batch and counting what it had to drop. A Collector
does not do that for you beyond the batch sizes above.

## Check it

Send a test span through the Collector, then look for the service on the Drops page: a refused
batch there with `invalid` outcomes means the Collector is forwarding something the door cannot
decode (usually `encoding: json` against a proto sender, or a path in `endpoint`).
