Let the kernel do the I/O.
Native io_uring for disk and network operations. Asynchronous I/O is part of the engine, from storage to the wire.
KEY-VALUE STORAGE FOR LARGE AI PAYLOADS
The whole payload.
A key to get it back.
Store inference state, tensor batches, and agent memory as large values. Thunder pairs core-local execution with encrypted storage and durable writes.
SELF-HOSTED · BUILT IN RUST · FREE TO RUN ↗
A serialized tensor can be hundreds of megabytes. A session snapshot can grow with every turn. Store the object your application works with, and retrieve it by key.
artifacts/encoder-v2/batch-009The complete serialized object.SHAPE + FORMAT VERSION + PAYLOAD BYTES
For 512 MiB values, configure a 4 GiB arena per reactor. Capacity depends on memory and workload. Sizing guide ↗
Work stays with the reactor that owns it. Each reactor has its own arena, shard range, and listening port. The client discovers the layout and routes keys directly.
Explore the architectureCRC32C(key) % 1024OWNING REACTOR01 Fixed shard layout.
Flexible reactor ownership.
02 Foreground execution.
Dedicated background work.
03 One starting address.
Client-discovered routing.
SMT ×2 means two hardware threads per physical core. Turn it off to explore a single-thread-per-core CPU. This local illustration uses Thunder’s key hash and uniform CPU topology; actual allocation follows CPU affinity. No data is sent to a database.
Four ways to use Thunder
in your AI application.
Store serialized inference state and prompt artifacts by key. Retrieve a previous result when your serving layer decides it can be reused.
inference/model-v3/prefix-8f2aVALUE → SERIALIZED STATEYour serving layer manages serialization, reuse, and transfer to the GPU.
Connect your applicationFrom the first model call to the next worker.
Explore the AI workflowsWritten in Rust. Built around local ownership,
asynchronous I/O, and durable writes.
Native io_uring for disk and network operations. Asynchronous I/O is part of the engine, from storage to the wire.
Each reactor serves its own shard range and memory arena. Work is assigned to an owner, with background maintenance alongside it.
Large payloads are first-class. Configure memory for your largest writes and measure the mixed workload your application actually runs.
Writes are journaled to disk before acknowledgement. TLS 1.3 protects the connection; every value is encrypted at rest.
Start with a verified round trip. Then measure your payload sizes, concurrency, and access patterns on the hardware you will actually run.
Evaluate your workloadSet a value. Get it back. The official client discovers the shard layout and routes requests to the right reactor from one starting address.
Connect your first clientClients for Linux, Windows, and macOS.
pip install thunderstrikefrom thunderstrike import Thunderstrike
db = Thunderstrike(
"127.0.0.1:7380",
tls_server_name="localhost",
tls_ca_pem=open("docker/tls/ca.crt").read(),
)
db.set(b"agent:42:state", b'{"step": 4}')
state = db.get(b"agent:42:state")Official client documentation THUNDERSTRIKE 0.1.6 / AVAILABLE ON DOCKER HUB
Run Thunder on your own hardware. Start with the sample, then put your real data through it.
Get started View the image on Docker Hub ↗LINUX / AMD64 + ARM64 / DOCKER COMPOSE
# Start your configured instance
$ docker compose up -d
# Check health, then verify the data
$ docker compose ps
$ docker compose run --rm sampleA FEW GOOD QUESTIONS
Thunder is a key-value database. It can hold serialized embeddings, documents, and other AI payloads by key. Similarity search, indexing, and model execution belong in your application or a dedicated retrieval system.
The deployment guide covers values of 512 MiB and more. For 512 MiB values, allocate a 4 GiB arena per reactor. A SET needs room for two copies of its value; your memory budget and workload determine practical capacity.
Thunder has 1,024 fixed logical shards. Each reactor owns a group of them, one arena, and one port. With two usable hardware threads per physical core, there is one reactor group per core. Without SMT, two physical cores supply one reactor group.
Use an official Thunder client for Python, Node.js, Go, Rust, .NET, Java, PHP, or C / C++. Thunder uses its own binary protocol over TLS 1.3, so Redis clients cannot connect directly. The official clients handle shard discovery and routing.
Docker with Compose v2, a Linux host kernel of 6.8 or newer, and a supported x86-64 AVX2 or ARM64 v8.2 CPU with Crypto Extensions. Plan at least 4 GiB of arena memory plus overhead for the default allocation. See the full requirements.
Thunder is proprietary software, distributed as thunderstrike and provided free of charge under its end user license. The license permits use as the database behind your own applications, subject to its terms.