← back to the storm

The Dry Stack

Database, queue and serializer in one process. Answers in microseconds.

Most backends are three systems taped together: a database to store data, a message queue to move it and a serializer to put it on the wire. Every seam costs a copy, a serialization pass and a network hop. The Dry Stack removes the seams. One process on one box ingests tens of millions of records a second, and you don't need a cluster to be fast.

Database store the data
copy, serialize, hop
Message Queue move the data
copy, serialize, hop
Serializer pack the data
three systems, glued at every seam
fused
DRY STACK
store
stream
serialize
one process, zero copies, no hops
8.7× DuckDB's ingest rate. We ran on 1 core, DuckDB got 32. 35× faster history replay than NATS JetStream, on the same box. 16 µs round-trip where gRPC takes 179 µs. Same machine, same messages. 0.87 ns to serialize a 16-byte message, with zero allocations.

Rivals get their fastest settings and every core. Medians from an open harness, and the losses are published too: Facts stores 3.8× more bytes per point than DuckDB, because it keeps columns raw for speed. Click a number for the full table.

Kill the leader yourself A two-node cluster is running right now. Cut the link, force a failover and watch it heal. A stale node can't win an election, so you never get two leaders. open the failover demo →
Platforms
Windows and Linux. SIMD-tuned for Intel, AMD and ARM, from a server rack down to a Raspberry Pi.
Languages
A C++ core with idiomatic C# on top. Native, zero-copy speed and no glue code to maintain.
Demos
Every demo below is a live engine answering real queries.
Facts
Time-series storage. SIMD queries over billions of records, inside your process.
Latest value in 0.2 µs. DuckDB takes 4.3 ms. Replaces InfluxDB or Timescale on one box.
Nio
Zero-copy networking. Tens of millions of messages a second on one connection.
206M msg/s across 64 connections on 12 threads, with no locks.
Zao
Serialization at memcpy speed. Tiered codegen picks the fastest path for each type.
1.4 to 3.6× ahead of MemoryPack. Protobuf is orders of magnitude behind.
plus Relay
Durable pub/sub built from the three. Nio moves the bytes and Facts holds the log, so the log is the broker.
About 4M msg/s persisted. JetStream manages 0.15M.
live demo → Benchmarks
The full query suite running on the box. Every API with its median and p99 latency, straight from the engine.
live demo → Market Wall
A command-center wall ingesting synthetic market data while queries hit the stream nonstop.
live demo → F1 Replay
One real Formula 1 race, ingested once and replayed. Point-in-time lookups and whole-race aggregates in microseconds.
live demo → Relay rides a failover
Kill the leader while a consumer reads. The counter keeps climbing with zero gaps, zero duplicates and no reconnect.
A durable subject is a replicated stream, so failover happens in the storage layer and the consumer never sees it.
live demo → Log Viewer
Scroll a huge live log. Only the visible window is read, including the unsealed tail.
live demo → Raspberry Pi
The same engine on a $35 ARM board, small enough to run where the data is born.
Code & usage
Real C# from the samples: open a store and query it, send messages, serialize a struct. Plus a look at the C++ core.
Setup & config
What it runs on, your first database in twenty lines, and every setting that matters with its default.
Technical overview (PDF)
Six printable pages with the full argument and every benchmark. Hand it to your favourite model and let it check the maths.

The PDF is A4. Also available in US Letter.

Early access

We're taking a few design partners. You get the stack and direct engineering help. We get your feedback and, if it works for you, a reference. Production licences are per deployment, simple and fair.

info@drylane.io
The Dry Stack is built by Drylane. More about Drylane.