Parseable
www.parseable.com- Category
- Developer Tools
- Rank
- No. 961Tools index
- Type
- TOOL
- Use case
- Deployment & Operations
- GitHub
- 2.6k stars
- Latest release
- v3.2.7
- Date
About
Parseable is an observability platform for logs, metrics and traces, written in Rust and built on object storage with open Parquet files. It is OpenTelemetry-native and supports SQL and PromQL queries. It can be deployed in a public or private cloud or used as a managed service.
What it does
A single Rust server that accepts logs, metrics and traces over HTTP, OpenTelemetry or an optional Kafka connector. Incoming events sit briefly on local disk in Arrow format, then get converted to compressed Parquet files and uploaded to Amazon S3, Azure Blob, Google Cloud Storage or a plain local directory. Queries run through Apache DataFusion, an embedded SQL engine, straight against those files. The same binary can run as one combined node or split into separate ingest and query nodes, and it bundles a web console, alerting, dashboards and role-based access control.
Why it's ranked here
The storage design is the case for it. By default buffered events are finished into Parquet every 60 seconds and pushed to object storage every 30, so telemetry lands in an open columnar format any Parquet reader can open, not a proprietary index. The server is AGPL 3.0, built on the Apache Arrow and DataFusion projects, and its large-file reads pin each object version so a concurrent rewrite causes a retry instead of a spliced, corrupt result. The open build has a real ceiling, though: two of its five server modes, indexing and Prism, print an enterprise notice and exit.
What's good
Query nodes read the same object-store files the ingest nodes write, so the two tiers scale separately. Large metadata objects are fetched as a 64 MiB first request followed by 16 MiB ranges, eight in flight, and the whole read restarts up to three times if the object changes mid-transfer. Queries run on a dedicated thread pool under a configurable timeout, and an optional memory check waits up to ten seconds before refusing a query as server busy rather than letting it exhaust the host. Release artifacts carry build-provenance attestations you can check with the GitHub command-line tool.
Tradeoffs
The quickstart logs in as admin with password admin, and once a basic-auth password checks out, that login is cached with no expiry until the password or roles change, so production use depends on the hardening guide the README links. AGPL 3.0 requires publishing your changes if you run a modified copy as a network service. Indexing and the Prism node type are paid features. Kafka ingest is a compile-time option, absent from the default build. The inter-node client can be configured to skip certificate checks. The recommended install pipes a remote script into the shell.
How to use it well
Fits teams already paying for S3, Azure Blob or Google Cloud Storage who want logs, metrics and traces kept as open Parquet that other engines can also query. Point OpenTelemetry exporters at it directly, start with the combined mode, and split ingest from query nodes once write load and read load diverge. Add a local hot-tier disk if dashboards keep scanning recent data; that option is refused with the local filesystem backend. Queries use the PostgreSQL flavour of SQL. It is a storage and query backend, so agents or collectors on each host still do the shipping.
Technical notes+
src/lib.rs sets LOCAL_SYNC_INTERVAL to 60s (Arrow staging files finished and compacted to Parquet) and STORAGE_UPLOAD_INTERVAL to 30s, and builds two reqwest clients with 3s connect and 30s request timeouts, HTTP/1.1 only; INTRA_CLUSTER_CLIENT calls danger_accept_invalid_certs when tls_skip_verify is set. src/main.rs dispatches on Mode: Query, Ingest and All start servers, Index and Prism print an enterprise message and exit(0); jemalloc is the global allocator off MSVC. src/storage/mod.rs implements get_object_ranged with RANGED_GET_PROBE_BYTES (64 MiB), RANGED_GET_CHUNK_BYTES (16 MiB), RANGED_GET_CONCURRENCY 8 and RANGED_GET_MAX_ATTEMPTS 3, pinning follow-up ranges with if_match on the first ETag and treating 412 or NotFound as an object change. src/query/mod.rs parses with PostgreSqlDialect, runs queries on a dedicated QUERY_RUNTIME under sql_timeout, and enough_available_memory polls cgroup or system memory for up to 10s before ServerBusy. src/option.rs defaults Parquet compression to Lz4Raw and bounds the event payload limit between 100 bytes and 100 MB. src/rbac/mod.rs gives sessions a one-hour EXPIRY_DURATION but tracks verified basic-auth keys with DateTime::MAX_UTC. src/handlers/http/modal/server.rs exposes tool listing and calling endpoints under an llm scope, gated by Action::Query. src/parseable/mod.rs selects local, S3, Blob or GCS storage from CLI subcommands and rejects a hot tier with the local store. Cargo.toml gates rdkafka and AWS MSK IAM signing behind the kafka feature.
Observed
- License
- AGPL-3.0-or-later (per source file headers)
- Language
- Rust, edition 2024
- Interfaces
- HTTP ingest and query API, OpenTelemetry ingest, bundled web console
- Storage backends
- Local filesystem, Amazon S3, Azure Blob Storage, Google Cloud Storage
- Query engine
- Apache DataFusion with the PostgreSQL SQL dialect
- On-disk format
- Apache Parquet in object storage, Arrow IPC for local staging
- Install surface
- Single binary via shell install script, PowerShell script on Windows
- Server modes
- All-in-one, ingest and query in the open build; index and Prism are enterprise only
- Optional build features
- Kafka connector behind a compile-time feature flag
- Build provenance
- Release artifacts attested, verifiable with gh attestation verify
- Test setup
- Integration-test and coverage CI workflows and docker compose test stacks in the repository tree
Read from README.md, Cargo.toml, src/main.rs, src/lib.rs, src/parseable/mod.rs, src/storage/mod.rs, src/query/mod.rs, src/event/mod.rs, src/handlers/http/modal/server.rs, src/otel.rs, src/option.rs, src/rbac/mod.rs.
Intel on Parseable
Tags
Tech Stack
Media

Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.