Ollygarden otel collector config decomposition

Decomposing an OTel Collector config

This skill is OllyGarden's opinion that a collector config should be decomposed only when the monolith has become painful — and left as one file when it hasn't. Decomposition buys smaller reviews, isolated testing, and clean per-environment variation, but it costs indirection: a reader now has to mentally merge several files to see the whole. When a config fits on a screen, that cost outweighs the benefit. Not splitting is a valid, common, and correct outcome.

The mechanism that makes decomposition work is the collector's deep merge of repeated --config file: sources (later keys win at each level), plus the file:/env:/yaml: providers. This skill owns the generic mechanics; the specialized ollygarden-otel-collector-k8s-daemonset skill is a consumer of this one for its own reference config.

It layers on upstream facts — point at these rather than duplicating them:

• Component config keys, defaults, signal support → the otel-collector skill.
• Behaviorally proving the decomposed config still does what it should (a filter still drops the right spans, a transform still sets the attribute) → the ollygarden-otel-collector-config-validation skill.

Default behavior: execute, but be critical first

When you're handed a collector config, assess it, then act directly — don't just advise. But the first decision is always whether to split at all:

1. Assess the config against the decision criteria below.
2. If it's not warranted → stop and say so. Make no changes. Report why the config is fine as one file (small, single pipeline, no env drift). A deliberate no-op is a success.
3. If it is warranted → split it, pick the best-fit strategy, write the files, and verify the merged result is behavior-equivalent to the original (see below).
4. Always report the reasoning — what you changed (or didn't) and why. The reasoning is part of the deliverable, not an afterthought.

When to decompose — and when not to

Decompose when two or more of these are true; a single weak signal is usually not enough:

• Size. The config has grown past a few hundred lines and reviews mean scrolling.
• Multiple signal pipelines. Traces, metrics, and logs pipelines with distinct receivers/processors that different people touch.
• Environment drift. Near-identical copies (prod/staging/dev) that have been copy-pasted and are drifting apart.
• Ownership boundaries. Distinct teams own distinct parts (platform owns receivers, observability owns processors, SRE owns exporter destinations).
• An oversized single block. One block dominates the file — a long tail_sampling policy list, a big set of Prometheus scrape jobs, a wall of OTTL statements.

Do NOT decompose when the config fits on a screen, has a single pipeline, has no environment variation, and is owned by one person. Indirection you don't need is a net negative. If in doubt and the signals are weak, leave it as one file and say why.

How to split

Pick the strategy that matches the config's actual organizational boundaries — don't apply a pattern mechanically. The three patterns, and how to choose between them, are in references/strategies.md:

• By signal pipeline — common.yaml + self-contained traces.yaml / metrics.yaml / logs.yaml. The default when signals are owned or iterated independently.
• By component type — base.yaml + receivers.yaml / processors.yaml / exporters.yaml. Fits when teams own pipeline stages.
• Environment overlays — a base plus production.yaml / staging.yaml / development.yaml that override values. Layers on top of either split above.
• Nested inclusion (${file:}) — pull one oversized, independently-ownable sub-block into its own bare-fragment file. Use sparingly.

Before writing files, read references/mechanics.md for the merge rules and the caveats that bite silently — most importantly that arrays are replaced, not merged (keep anything that varies together, especially a pipeline's processors: list, in one file) and that ${file:} paths resolve relative to the collector's working directory, not the including file.

Verify before you're done

A decomposition that changes behavior is a bug, not a refactor. Never hand back a split you haven't merged and checked. Validate the merged set (never a lone fragment) and diff the fully-resolved output against the original monolith. The commands and the behavior-equivalence check are in references/verifying.md. For proving a specific processor/connector still transforms telemetry correctly, hand off to the ollygarden-otel-collector-config-validation skill.

Report template

End every run with a short report so the reasoning ships with the change:

Decision: <decomposed | left as one file>
Why: <which criteria were / weren't met>
Strategy: <by-signal | by-component | overlays | nested — or n/a>
Files: <the file set produced, or "unchanged">
Verification: <merged validate result + equivalence check, or "n/a">

Pitfalls

• Splitting a config that didn't need it. The most common mistake. Indirection has a cost; a screen-sized single-pipeline config should stay one file.
• Handing back an unverified split. If you didn't validate the merged set and confirm equivalence, you don't know the refactor was behavior-preserving.
• Splitting a pipeline's processors: array across files. Arrays are replaced, not merged — the last file wins and silently drops the rest of the list. Keep it in one file.
• Applying a pattern that doesn't match ownership. Splitting by signal when one team owns everything just adds files to merge in your head. Match the split to real boundaries.
• Over-using ${file:} nested inclusion. Externalize only blocks that are both large and independently ownable; each include is another file to open.

Cross-references

• Component facts (keys, defaults, signal support): otel-collector.
• Behavioral proof that the split config still transforms telemetry correctly: ollygarden-otel-collector-config-validation.
• A worked application of these mechanics on a real reference config: ollygarden-otel-collector-k8s-daemonset.