Open source observability: which platform to pick

Five open source tools take in logs, metrics and traces and keep them on your own servers, but each one stores them differently. SigNoz and Superlog write everything to ClickHouse, and Superlog also hands Claude the job of grouping errors into incidents. Traceway uses ClickHouse as well in server mode, yet also fits in a single container on SQLite or DuckDB. OpenObserve runs as a single binary and keeps its data as Parquet files, on local disk or in an S3 bucket. Netdata starts from the machines: an agent on each one samples metrics every second and stores them locally.

An OpenTelemetry backend, or an agent on every machine

SigNoz, OpenObserve, Traceway and Superlog are OpenTelemetry backends: you instrument your applications with an OpenTelemetry SDK, and they send traces, logs and metrics over OTLP. Traceway takes OTLP/HTTP straight from the SDK, with no Collector in between. SigNoz ships its own OpenTelemetry Collector as part of its stack and also accepts Prometheus metrics. Superlog receives OTLP through an ingest proxy and a collector, and its Vercel, Railway and Render connectors forward logs and metrics without any OTLP wiring on your side; AWS and Cloudflare have connectors too.

Browsers and mobile apps need more than OTLP. OpenObserve ingests RUM data from browsers and mobile apps. Traceway ships its own SDKs for React, Vue, Svelte, Flutter, Android, iOS and React Native, with session replay on the web and in Flutter.

Netdata comes at the problem from the other side. Its agent, installed on each machine, finds containers, systemd services and processes on its own and samples their metrics every second, with no application instrumentation. It has accepted OpenTelemetry logs since v2.11.0 in August 2026, and traces since v2.12.0, as an early-access feature. Reading them takes a Netdata Cloud account, free or paid: Netdata's access-control docs block the systemd journal, Windows events, the process list and network connections for anyone who isn't signed in. Metrics, alerts and anomaly detection work without an account.

Probing a service from outside and posting its state on a status page is a job for a different kind of tool, covered in the comparison of open source status pages and uptime monitors; OneUptime is one of them and takes in logs, metrics and traces over OpenTelemetry too.

ClickHouse, object storage or an embedded database

ClickHouse, a columnar database, holds the telemetry for three of the five tools, and never on its own. SigNoz starts ClickHouse, ClickHouse Keeper, PostgreSQL for metadata, its collector and its backend, and asks for at least 4 GB of memory for "a small test installation". Superlog runs PostgreSQL 16, ClickHouse 26.1 and a collector, plus four Node apps (web UI, API, ingest proxy and worker), and its README only covers a development setup. Traceway's Compose file pairs ClickHouse 24.8 with PostgreSQL 17.

Traceway can also do without ClickHouse. Its SQLite image fits in one container, but the docs limit it to trying Traceway out or running a small instance. The DuckDB image is the recommended single-container option, and the same docs say it keeps dashboards responsive with 10 to 100 times more rows. Both run as a single instance: more than one backend requires PostgreSQL and ClickHouse. DuckDB budgets about 80% of the RAM it believes is available, and in a memory-capped container without DUCKDB_MEMORY_LIMIT, the backend gets OOM-killed under ingest.

OpenObserve needs no external database on a single instance: metadata goes to SQLite, data to Parquet files on local disk. Object storage (S3, GCS, MinIO, Azure Blob...) is switched on with ZO_LOCAL_MODE_STORAGE and becomes mandatory for high availability, which runs on Kubernetes with Helm and adds PostgreSQL or MySQL for metadata and NATS to coordinate the nodes. The README claims storage 140 times cheaper than Elasticsearch, which it credits to Parquet compression and the price of object storage. Reads are paid for in cache: by default a querier gives half its memory to Parquet files.

Netdata has no central database. Each agent writes to its own storage engine in three tiers: per-second data for 14 days, per-minute for 3 months and per-hour for 2 years, each tier capped at 1 GiB. "Parent" agents receive data from the others and hold their history, and Netdata Cloud queries agents and parents without centralising storage.

What each tool needs to run

Read from the repositories and documentation on 7 October 2026.

Few of these projects say what resources a given ingest volume calls for. SigNoz's production sizing adds up to 56 cores and 152 GiB (three collectors, two ClickHouse shards, three ZooKeeper nodes, PostgreSQL and the backend) without saying what volume it is meant for. OpenObserve's capacity planning page points to an external spreadsheet, Traceway gives no CPU or memory figures for its ClickHouse mode, and Superlog documents none at all. Netdata puts numbers on its agent: about 150 MiB of RAM and 5% CPU in a default production setup, and 100 MiB with machine learning and alerts off and ephemeral storage.

OpenObserve keeps everything for ten years by default: unless you set ZO_COMPACT_DATA_RETENTION_DAYS, nothing is deleted for 3,650 days. Its data is also immutable, and only whole retention periods get dropped. SigNoz keeps logs and traces for 15 days and metrics for 30, and a new retention setting only applies to data ingested afterwards. Superlog sets no retention on the tables its collector writes to, only a 30-day one on the tables derived from traces. On Netdata, each tier's quota is shared by every agent that streams to the same parent.

Where the projects stand

SigNoz dates back to January 2021 and has never tagged a 1.0: thirteen minor versions shipped between 0.133.0 on 15 July and 0.145.0 on 5 October 2026, roughly one a week. Its install.sh script and Docker Compose files are deprecated in favour of Foundry, a separate installer under AGPL-3.0. OpenObserve, whose repository dates from February 2023, shipped v1.0.0, its first generally available 1.x release, on 11 September 2026, and has published 43 stable releases in 2026.

Traceway is the youngest: repository created in December 2025, 89 backend releases between May and October 2026, and one developer behind 823 of the 1,056 commits on the main branch. Its 2.0, released on 21 September, is a breaking upgrade: it changes the trace tables and part of the API, and history is copied over after a backup with V2_MOVE_OVER=true.

Superlog has published no release or tag, and nothing has been merged into its main branch since 14 September 2026, with 69 pull requests still open; one developer wrote 468 of its 488 commits. On 1 October its docs were rewritten: they now name a different repository from the same company, responder-oss, as Superlog's open source application, and say the hosted service runs that code.

Netdata shipped five minor versions between November 2025 and September 2026. The v2.12.0 notes open with a warning about random crashes on startup and recommend v2.11.1 until v2.12.1 is out. Its kickstart.sh installer defaults to the nightly channel with daily automatic updates; --stable-channel sticks to stable releases.

Licences, enterprise editions and cloud plans

Traceway says it holds nothing back. Its README claims MIT with no BSL and no "open core", its website promises full feature parity with Traceway Cloud, and SSO through Google, GitHub or OIDC is part of the open source edition. The cloud charges for hosting: free up to 1 GB of ingest a month, then 24.99 dollars a month for 150 GB.

SigNoz is under MIT except for its ee/ and cmd/enterprise/ directories, which fall under the SigNoz Enterprise licence and need a subscription in production. The pricing page leaves SSO in the Community edition but keeps SAML, anomaly detection, multiple ingestion keys and one-click AWS integrations for paid plans. Noz, the built-in AI assistant, only exists on SigNoz Cloud, whose Teams plan starts at 49 dollars a month, with logs and traces billed at 0.30 dollars per GB. Enterprise, also available self-hosted with support, starts at 4,000 dollars a month.

OpenObserve is under AGPL-3.0 or later: anyone who serves a modified version over a network must publish its source. The open source edition has no SSO and no roles, and every user has access to every feature. The enterprise edition, a separate image under a commercial licence, adds SSO (OIDC, SAML 2.0, LDAP/AD), RBAC, an audit trail, federated search across clusters and automatic redaction of sensitive data. Self-hosted, it is free up to 50 GB ingested per day.

Superlog publishes its community edition under Apache-2.0, and its README calls the project open-core. The runner behind real investigations is an external module the repository doesn't include; the bundled one calls no model and proposes no fix. Error grouping, on the other hand, lives in the repo and calls Claude, with Anthropic as the only provider wired in. To classify an error, it sends the exception's title, message and stack trace, its attributes and trace IDs, plus the title and the start of the error message of every open incident. Without an API key, an error the fixed rules can't place opens an incident of its own.

Netdata licenses its agent under GPL v3 or later, but the web UI is closed source: its NCUL1 licence forbids modifying it or using it with anything other than Netdata agents and Netdata Cloud. Without a subscription, dashboard access stops at five nodes, and configuring collectors from the UI, RBAC and SSO are paid features. The pricing pages list Business at 4.50 dollars per node per month billed annually, and Homelab at 90 dollars a year for non-commercial use.