Back to catalog
Grafana Tempo
AGPL-3.0Grafana Tempo replaces costly APM tracing SaaS like Honeycomb by streaming OpenTelemetry spans into cheap object storage using Apache Parquet.
Open-source alternative to — compare it against:
Overview
Commercial tracing platforms like Honeycomb price by event volume, forcing teams to discard most spans through aggressive sampling. Grafana Tempo strips out dedicated index databases and writes raw traces straight to object storage in Apache Parquet format. By eliminating per-event SaaS fees and database maintenance, it allows organizations to retain complete request telemetry on their own infrastructure.
Why Grafana Tempo instead of Honeycomb
Honeycomb delivers sub-second queries across high-cardinality attributes using an in-memory columnar engine, but that speed comes chained to event pricing. When distributed systems generate millions of spans hourly, teams must drop ninety percent or more of their traces to control bills. Grafana Tempo prioritizes write throughput over instant indexing. By streaming spans directly into columnar Apache Parquet files on object storage and buffering writes through Kafka with single replication, Tempo sustains high ingestion volumes with minimal compute overhead. Teams retain all traces rather than gambling on head-based sampling.
The performance trade-off appears during retrieval. Honeycomb evaluates arbitrary exploratory queries almost instantly, whereas Tempo executes TraceQL by fanning out worker processes across object storage blocks. Fetching a trace by identifier or jumping from a Prometheus exemplar takes milliseconds, but broad queries across wide windows require streaming Parquet blocks over the network and consume noticeable CPU. For teams already running Prometheus and Loki, Tempo provides predictable operational costs and ample ingestion headroom in exchange for slower deep scans.
What it does well
- Object storage persistence: Writes traces directly to AWS S3, Google Cloud Storage, Azure Blob, or MinIO, eliminating Cassandra or Elasticsearch clusters.
- Columnar Parquet format: Stores span attributes in standard Apache Parquet blocks, enabling external analytical tools and data lakes to query traces directly.
- TraceQL query language: Allows developers to filter traces by duration, status code, structural relationships, and arbitrary tags using syntax familiar to LogQL users.
- Built-in metrics generator: Derives RED metrics and service dependency graphs directly from incoming span streams and pushes them to Prometheus.
- Multi-protocol ingestion: Ingests OpenTelemetry, Jaeger, Zipkin, and Kafka streams natively without requiring translation proxies.