> 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/scaling.md).

# Scaling Profiles

Hermes shares a local / standard / scaled vocabulary, but the achievable tier depends on how Salesforce delivers each telemetry kind, not on Hermes compute.

Every Hermes runtime shares the same three-tier vocabulary — `local`, `standard`, `scaled` — but not every runtime can reach every tier. The reason isn't Hermes compute; it's the **Salesforce ingestion mechanism** each runtime depends on. Iris and Argus RTEM are both bound by a single-active-subscriber limit on their Pub/Sub stream; Argus ELF pulls files over REST instead, and files can be sharded across workers. That structural difference, not raw throughput, is what caps each runtime's ceiling.

***

## Runtime Matrix

| Runtime        | Ingestion mechanism            | Concurrency model                        | Achievable profiles           |
| -------------- | ------------------------------ | ---------------------------------------- | ----------------------------- |
| **Iris**       | Pub/Sub Platform Events        | Single-active subscriber per org/topic   | `local`, `standard`           |
| **Argus RTEM** | Pub/Sub Platform Events (RTEM) | Single-active subscriber per org/channel | `local`, `standard`           |
| **Argus ELF**  | Bulk `EventLogFile` REST       | File-sharding + worker leases            | `local`, `standard`, `scaled` |

Argus ELF is the **only horizontally-scalable Hermes runtime** — because its unit of work is a file claimed by lease, adding workers adds throughput. Iris and RTEM are each capped at one active subscriber per stream; you scale their headroom by tuning that one process, not by adding more of them.

***

## What Each Tier Means

* **`local`** — a dev/single-process deployment. State is ephemeral (a `file`-backed store, or none), and data loss on restart is acceptable. This is how you run Hermes to see your first trace, not how you run it in production.
* **`standard`** — a production deployment with managed state — DynamoDB on AWS, Azure Table Storage on Azure — behind a single active process per subscription. This is the tier Iris and RTEM top out at, and where most ELF deployments also start.
* **`scaled`** — a larger, bursty, multi-org production deployment. Only Argus ELF reaches this tier, because it's the only runtime where adding capacity means adding workers rather than hitting the single-active ceiling.

***

## Sizing ELF: The Load Unit

Because Argus ELF pulls files instead of subscribing to a stream, its capacity is sized by data volume, not event rate. Hermes expresses ELF sizing in **load units**:

**1 load unit = 1 GB/hour sustained = 24 GB/day**

This unit applies to **Argus ELF only**. Iris and Argus RTEM aren't sized by data volume at all — they're sized by event rate and stream count, since their ceiling is the single-active-subscriber limit described above, not throughput. The formulas that turn load units into worker counts are in [Sizing & Cost](https://docs.pharos.ai/open-telemetry/hermes/sizing-and-cost.md).

***

## Per-Worker ELF Throughput

Per-worker decode throughput (decode → mapping → identity → dedup, no export) was measured with a Go benchmark on an Apple M1 Max workstation on 2026-05-30 at about 64 MB/s; the Graviton (≈24 MB/s) and x86 Fargate (≈13 MB/s) planning figures are derived from per-core ratios of 1/2.7 and 1/5.

Rule of thumb for sizing the `scaled` tier: `workers ≈ ceil(target sustained MB/s ÷ per-worker rate)`.

ELF's autoscaler target-tracks the `elf_lag_seconds` gauge (published as `argus_elf_lag_seconds`), aiming for roughly **≈ 60s** of lag on `scaled` and **≈ 300s** on `standard`.

***

## "Trial/POC" Is a Posture, Not a Profile

Running Hermes for a trial or proof-of-concept doesn't require its own scaling tier. A trial deployment typically runs at `local` or `standard`, sized down — it's a **deployment posture**, a decision about how much infrastructure to stand up for how long, not a distinct scaling profile with its own ceiling. The same `local` / `standard` / `scaled` vocabulary applies whether the deployment is a two-week trial or a permanent production install.

***

## 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>📐 Sizing &#x26; Cost</strong></td><td>Starting compute shapes, load-unit formulas, and the AWS/Azure cost model</td><td><a href="https://docs.pharos.ai/open-telemetry/hermes/sizing-and-cost.md">Sizing &amp; Cost</a></td></tr><tr><td><strong>📑 EventLogFile</strong></td><td>The only Argus runtime that reaches the scaled tier</td><td><a href="https://docs.pharos.ai/open-telemetry/hermes/argus/elf.md">EventLogFile</a></td></tr><tr><td><strong>⚙️ Configuration</strong></td><td>Deployment sizing and state-store setup per cloud</td><td><a href="https://docs.pharos.ai/open-telemetry/configuration.md">Configuration</a></td></tr></tbody></table>
