Methodology

What this site tracks

OSS driver / SDK / agent download volumes for SaaS observability vendors, as a directional leading indicator of paid managed-cloud consumption.

This is a research instrument, not a forecast tool. Every claim should read as "signal moved in this direction", never "this implies X% ARR growth".

The "SDK downloads → ARR, ~6-month lag, ~98% correlation" finding holds primarily for API-first vendors (OpenAI, Anthropic) where SDK calls drive billing directly. For SaaS infra vendors with self-hosted free tiers, free-quota cloud, and compatibility-layer leakage, the linkage is much weaker. Site copy uses directional language only.

2. Signal-quality tagging at package level

Pure — downloads ≈ adoption of the vendor's revenue-bearing path.
Leaky — meaningful free / self-host / compatibility-layer / proxy use.
Indirect — package is not first-party to the vendor; used as a proxy or meta-signal.

A driver's downloads growing in isolation says little. Within-sub-segment market share is the defensible analysis.

4. SDK-as-signal validity decays by sub-segment

Sub-segment	SDK signal strength	Reason
APM	🟢 strong	SDK install is the obligatory path
RUM	🟢 strong	Front-end SDK is required
Infrastructure monitoring	🟡 medium	Agent binaries dominate
Log management	🔴 weak	Forwarders dominate, not SDKs
SIEM	🔴 very weak	Connector-driven
LLM observability	🔴 weak	Multiple onboarding paths (proxy, framework auto-tracing, OTel exporter)

This MVP tracks APM and LLM Obs only. For LLM Obs, packages are selectively included: only those whose SDK install genuinely reflects adoption (Phoenix self-host, Braintrust eval-driven, OpenLLMetry meta). Helicone (proxy-first) and LangSmith (transitive dep of LangChain) are deliberately excluded.

5. The methodology page is canonical

Read this page before drawing conclusions from a leaderboard row.

What signals do not capture

Self-hosted use — paid cloud doesn't see this; SDK does.
Free-tier use — registered but under quota.
Proxy onboarding (LLM Obs) — when the user changes base_url to route through the vendor without installing the vendor's SDK.
Compatibility leakage — pymongo downloads include DocumentDB and Cosmos DB users; elasticsearch includes OpenSearch.
Self-built observability — internal stacks at FAANG-scale companies. Use OTel as a directional proxy.

Data sources

Source	Window	Auth
pypistats.org	last 180 days	none
api.npmjs.org	full history	none
BigQuery `pypi.file_downloads` (deferred to v1.1)	full history	GCP credential

PyPI's 180-day API limit means PyPI packages show only ~6 months of history at MVP launch, while npm packages show ~5 years. v1.1 closes the gap via BigQuery.

Per-package caveats

Excluded LLM Obs vendors (and why)

Helicone — primary onboarding is proxy mode (changing base_url of the OpenAI client). The Helicone SDK is optional. SDK downloads systematically understate true adoption.
LangSmith — langsmith is a transitive dependency of langchain. SDK downloads track LangChain installs, not LangSmith activation.
DDOG / NR / DT LLM Observability product extensions — LLM-specific functionality is shipped inside existing APM SDKs (ddtrace, newrelic, etc.) with no distinct package.