> For the complete documentation index, see [llms.txt](https://docs.pharos.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.pharos.ai/open-telemetry/hermes/backends.md).

# Backends & Routing

Backends are OTLP routes behind one shared exporter — direct to a vendor endpoint, or through a Collector for central sampling, filtering, and fan-out.

Hermes treats every OTel backend as an interchangeable OTLP route behind **one shared exporter** — there's no per-backend fork of the pipeline. Switching backends, or sending to more than one, is a routing decision, not a code change.

***

## Direct Mode vs. Collector Mode

The export destination is declared once, at the top level of [`ORG_CONFIGS`](https://docs.pharos.ai/open-telemetry/configuration/org-configs.md), as exactly one backend block:

* **Direct mode** — a `honeycomb`, `splunk`, or generic `otlp` block: Hermes talks straight to the vendor's OTLP/gRPC endpoint, carrying that vendor's API key or token in its export config.
* **Collector mode** — a `collector` block: Hermes exports to an OpenTelemetry Collector instead, whether a sidecar (`http://127.0.0.1:4317`, the default) or a service reachable by DNS. The Collector then forwards on to the real backend.

Collector mode exists for more than indirection: it lets you apply **central sampling, filtering, and egress control** in one place rather than per-runtime, **fan out** the same telemetry to more than one destination, and keep backend tokens **out of `ORG_CONFIGS`** entirely — the Collector holds the credential, not Hermes. A single Hermes org exports to a single endpoint; fan-out is the Collector's job.

***

## What Each Backend Carries

| Backend                                                                                   | Direct block                                                      | Through a Collector                                                           |
| ----------------------------------------------------------------------------------------- | ----------------------------------------------------------------- | ----------------------------------------------------------------------------- |
| **Honeycomb**                                                                             | `honeycomb` — traces, metrics, and logs                           | All three                                                                     |
| **Splunk Observability Cloud**                                                            | `splunk` — **traces only**; `signals` resolves to `traces`        | Traces and metrics; Salesforce log records via Splunk's log ingestion path    |
| **Datadog**                                                                               | None — Datadog's OTLP intake is HTTP-only and Hermes exports gRPC | All three, through the Collector's Datadog exporter                           |
| **Grafana Cloud / self-hosted LGTM**                                                      | None — same reason                                                | All three, through Alloy or a Collector fanning out to Tempo, Mimir, and Loki |
| **Any other OTLP/gRPC endpoint** (New Relic, Dynatrace, LogScale, a self-hosted receiver) | `otlp` — you supply the endpoint and auth headers                 | All three                                                                     |
| **Local JSONL file**                                                                      | `file` — local development and debugging only                     | —                                                                             |

Elasticsearch/OpenSearch is reached like any other OTLP target: through a Collector with the vendor's exporter.

![Salesforce telemetry as OpenTelemetry traces in a Grafana Tempo backend.](https://4264222071-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2Fdl9tmJ56lvwg5XY63QYr%2Fuploads%2Fgit-blob-0cff67f730901cfebcb29842b1807c932c629a22%2Fgrafana-traces-drilldown.png?alt=media)

***

## Metric Temporality

Hermes exports its counter/sum metrics (`argus.*.count`, `argus.elf.*.count`) using OTLP **delta** temporality. Honeycomb, Splunk, and Datadog all accept delta sums natively — no extra configuration needed.

{% hint style="warning" %}
**Grafana Cloud Mimir and self-hosted Prometheus/Mimir silently drop delta sums.** Counters simply won't appear — gauges, spans, and logs are unaffected. Add a `deltatocumulative` processor to the Collector's Prometheus-facing metrics pipeline so counters convert to the cumulative temporality Prometheus expects. See [Grafana Cloud & LGTM](https://docs.pharos.ai/open-telemetry/configuration/backends/grafana.md) for the processor wiring.
{% endhint %}

Separately, any Prometheus-family backend receiving Argus **ELF** metrics needs `out_of_order_time_window` raised to **72 hours or more** — ELF metrics carry historical event timestamps rather than current wall-clock time, and Prometheus's default out-of-order window is far too narrow to accept them. RTEM metrics use current wall-clock timestamps and aren't affected by this setting.

***

## Signal-Visibility Caveat

Not every signal type is visible in every backend's UI. Some Argus event types are **no-span** — they emit a log, or a log and metric, with no span at all (see [Architecture](https://docs.pharos.ai/open-telemetry/hermes/architecture.md)). A backend whose dashboards are built around span counts, like Honeycomb, won't surface those events on its main boards, because there's no span to count.

That telemetry isn't missing — it's just visible somewhere else. A no-span Argus event still shows up in a log-oriented view (for example, Splunk's Log Observer) or a metric-oriented view (for example, Splunk's Metric Finder). Where to look depends on which signal that particular event type actually emits, not on where you'd expect to see "everything."

***

Backend routing here covers the concept — which modes exist and what each backend carries. The per-backend configuration detail — API keys, endpoint formats, Collector wiring, queries, dashboards, troubleshooting — lives in [Backend Configuration](https://docs.pharos.ai/open-telemetry/configuration/backends.md) and its per-backend runbooks.

***

## Related

<table data-view="cards"><thead><tr><th></th><th></th><th data-hidden data-card-target data-type="content-ref"></th></tr></thead><tbody><tr><td><strong>🎯 Backend Configuration</strong></td><td>Backend blocks, endpoint rules, and Collector wiring, with per-backend runbooks</td><td><a href="https://docs.pharos.ai/open-telemetry/configuration/backends.md">Backend Configuration</a></td></tr><tr><td><strong>🔗 Cross-Surface Correlation</strong></td><td>How Iris and Argus traces stay stitched together before export</td><td><a href="https://docs.pharos.ai/open-telemetry/hermes/correlation.md">Cross-Surface Correlation</a></td></tr></tbody></table>
