GreptimeDB v1.3.0-beta.2 performance numbers in the time-series market
GreptimeDB achieves 2M rows per second using gRPC Bulk protocol, outperforming MySQL and PostgreSQL ingestion rates by up to 55x. The database also demonstrates superior storage efficiency, using 18x less storage than TimescaleDB and 1/11 the data volume of InfluxDB.
Ingestion throughput and the 55x gap
GreptimeDB hits 2M rows per second when using the gRPC Bulk protocol with Arrow Flight. This rate exceeds the 1.45M rows per second from gRPC Stream and the 1.19M rows per second from the gRPC SDK. In comparison, MySQL and PostgreSQL insertions reach only 36K to 69K rows per second because of SQL parsing and synchronous connections. The throughput gap between the fastest and slowest ingestion methods reaches 55x, as the 2M rows per second achieved via gRPC Bulk significantly exceeds the 36K to 69K rows per second produced by MySQL. OTLP Logs and InfluxDB LP protocols achieve approximately 1M rows per second. High series cardinality with 1M series reduces gRPC Bulk speed by 32%, but the append-only model of OTLP Logs skips primary key maintenance and remains largely unaffected. A test using 10M rows and 1M time series with 5 workers showed that increasing batch size nearly doubles gRPC Bulk throughput from 1.14M to 2.06M. The system performs near zero-copy writes because it uses Arrow internally.
Comparison of storage efficiency and scalability
GreptimeDB writes 2.17x faster than TimescaleDB on the standard TSBS benchmark and uses 18x less storage. In head-to-head benchmarks, GreptimeDB ingests 4.7x faster than InfluxDB and occupies about 1/11 the data volume. For teams evaluating Elasticsearch as a log alternative, GreptimeDB ingests 4.7x faster and stores 1/11 the data while using a fraction of the memory required by the legacy system. TimescaleDB manages 100K to 500K inserts per second on a single node, but it hits a write ceiling because it inherits the single-primary limitation of PostgreSQL. A fleet of 10,000 IoT sensors reporting every second generates 864 million rows per day. To avoid performance drops, the documentation recommends sizing TimescaleDB chunks so the most recent chunk and its indexes stay within 25% of available RAM. InfluxDB 3.0 uses a Rust-based engine with Apache Arrow and DataFusion, but it cannot join with other data sources as easily as TimescaleDB. ClickHouse processes billions of rows per second for COUNT, SUM, and AVG operations and delivers 10-100x faster analytical queries than PostgreSQL on identical hardware, yet it lacks ACID transactions and makes frequent corrections to historical data difficult. ClickHouse compresses a 1TB raw dataset to 100-200GB, while columnar storage in general typically reaches 90-95% compression ratios. GreptimeDB replaces the 5-8 component stack of Prometheus, Thanos, or Mimir with one PromQL-compatible, object-storage-native database. It also acts as a Grafana Loki alternative with full-text and skipping indexes, and it has proven effective at 300 TB across 80+ clusters in production. The cloud-native time-series database market expects a 5.6% CAGR through 2033, driven by the need for real-time analytics in IoT and financial sectors. You should decide if your team needs the simplicity of a single-node architecture or the horizontal scalability of a distributed system.
Query performance in the October beta release
The v1.3.0-beta.2 release arrived on October 15, 2026, and enables distributed write retries by default. These retries use exponential backoff and jitter when the frontend encounters busy regions or unavailable datanodes. The update improves 30-day PromQL range queries over node metrics with a median speed increase of 1.43x. SQL queries over 100M agent observability records show a median improvement of 1.20x. The release also fixes errors where PromQL comparisons and min/max functions returned incorrect results for NaN ratios or empty traffic. Fulltext bloom indexes no longer skip matching rows because the index now splits text where the matches_term function does. The new Signals Explore view provides a drilldown capability for metrics, logs, and traces, and it lets users apply label filters to the aggregated data. The release also fixes a subquery issue where count_over_time(x[5m:]) returned 300 instead of the expected 5. The updates also fix an issue where binary operations on folded histograms could drop matching series when query parallelism was above 1. The updates also resolve a case-sensitive LIKE pruning issue, though existing SSTs require a manual index rebuild. The Table Semantic Layer records the signal, source, and unit behind each table to improve data organization. The roadmap for the fourth quarter includes adaptive resource management in its second phase and production-ready remote compaction. Will the planned adaptive resource management in the fourth quarter resolve the remaining memory pressure in high-concurrency environments?