slap_kv provides a partitioned key-value store, with data stored durably in object storage.

slap_kv is built on slap_cluster and slap_slatedb. Each row is addressed by a partition and a key. A partition's rows are stored together, in key order, in one of a fixed number of SlateDB databases, called shards. Writes are acknowledged once they are durable, reads return only durable rows, and a write can be made conditional on a row's version.

Use slap_kv when you need records grouped by a partition, such as a user's settings or a document's metadata. It routes each partition to a shard, supports listing keys in order, and uses versions to detect concurrent updates. This saves you from building shard ownership and conditional writes on top of slap_slatedb. If you need direct SlateDB transactions or a different key layout, use slap_slatedb instead. slap_kv does not support cross-partition scans or cross-partition transactions.

slap_kv runs inside your application, on one node or several. You start Slap.KV.Cluster, call Slap.KV in-process, and can mount Slap.KV.HTTP.Router in your Plug pipeline. A call for a partition goes to the node that owns its shard. The slap package's standalone server script serves slap_kv over HTTP.

About Slap

This package is part of Slap.

Slap provides Elixir bindings for SlateDB (an embedded key-value engine optimized for storing data in object storage), as well as services built on top of these bindings. The services run inside your application, on one node or several.

Services:

  • slap_streams: a Durable Streams server, with HTTP, long polling, and SSE
  • slap_kv: partitioned key-value storage with conditional writes, ordered scans, and an HTTP API
  • slap_files: file storage with metadata and conditional replacement; small files are kept in slap_kv records, which saves object-store PUTs
  • slap_yjs: Yjs documents for collaborative editing, with replication and presence across nodes

Building blocks:

  • slap_slatedb: the SlateDB bindings
  • slap_cluster: SlateDB databases run as shards across nodes, with placement, routing, and failover
  • slap_snapshot_log: an append-only log of changes that can be condensed into snapshots

All-in-one:

  • slap: all Slap server libraries, and Slap, which starts the services you select under one supervisor

Inspection:

  • slap_inspect: a read-only TUI and CLI for browsing Slap data; installed separately

The slap package pulls in all seven server libraries. If you only need a subset, add those packages directly.

Installation

Requires Elixir 1.18 or later.

To depend on slap_kv directly, add it to the dependencies in your application's mix.exs:

defp deps do
  [
    {:slap_kv, "~> 0.6.1"}
  ]
end

Run mix deps.get. API documentation is on HexDocs.

Example

In an application that depends on slap_kv, run this in iex -S mix with a fresh local directory:

iex> {:ok, _} = Slap.KV.Cluster.start_link(store: {:local, "/tmp/slap-kv"}, path: "kv", shards: 1); :ok
:ok

iex> {:ok, v1} = Slap.KV.put("user:42", "profile", "Ada", if_version: :absent); :ok
:ok
iex> {:ok, %{value: value, version: ^v1}} = Slap.KV.get("user:42", "profile"); value
"Ada"
iex> {:error, {:conflict, ^v1}} = Slap.KV.put("user:42", "profile", "other", if_version: :absent); :conflict
:conflict
iex> {:ok, v2} = Slap.KV.put("user:42", "profile", "Ada Lovelace", if_version: v1); :ok
:ok
iex> {:ok, %{rows: rows, cursor: nil}} = Slap.KV.scan("user:42", prefix: "pro"); rows
[{"profile", "Ada Lovelace"}]
iex> :ok = Slap.KV.delete("user:42", "profile", if_version: v2)
:ok

For local storage, see choosing the shard count; before running a script against a store an application is using, see running another cluster on the same store.

An application can define its own cluster module and pass it to KV calls and the HTTP router:

defmodule MyApp.KVCluster do
  use Slap.KV.Cluster, otp_app: :my_app
end

children = [{MyApp.KVCluster, store: {:local, "/tmp/slap-kv"}, shards: 1}]

Add these children to your supervision tree. Once it is running, pass the cluster module to each operation:

{:ok, _} = Slap.KV.put("users", "42", "Ada", cluster: MyApp.KVCluster)

Supervise one KV cluster per VM. A Streams cluster can run on the same VM. Run one Slap.Cluster.Peers per VM; its connections to other nodes serve every cluster on that VM.

A failed store probe normally prevents the cluster and its supervising application from starting. With probe: :background, the cluster starts without shards and retries the probe. Operations return {:error, :unavailable} until the shards open.

Before boot work that needs every shard, call Slap.KV.Cluster.await_ready(30_000) and handle {:error, :timeout}. Slap.KV.Cluster.status() reports current readiness and the latest failures without accessing storage; see cluster readiness.

Flush interval

A durable write waits for SlateDB to flush its write-ahead log (WAL) to object storage. Slap.KV.Cluster sets flush_interval to 10 ms; SlateDB's default is 100 ms. Pass settings: %{flush_interval: "50ms"} to change it. Other settings leave the 10 ms default intact.

Lower intervals reduce the wait for acknowledgements but increase object-store PUTs. Each shard with pending writes can issue up to one WAL PUT per interval; idle shards issue none. Choose the interval for the expected write rate, shard count, and store pricing. For :memory and {:local, dir} tests, settings: %{flush_interval: "2ms"} reduces the wait without request charges.

Semantics

  • Versions. A row's version is the SlateDB sequence number of the write that stored it, so it changes with every write and only increases within a shard. put/4 and delete/3 accept if_version: version, and put/4 also accepts if_version: :absent. When the condition fails, the call returns {:error, {:conflict, current}}, where current is the row's current version, or nil if the row does not exist.
  • Durability. A write returns once SlateDB reports it durable. Reads (get/3, scan/2) use SlateDB's :remote durability: they see only durable rows, so they never return a write that could still be lost.
  • Uncertain writes. After an :unavailable or :timeout reply, the write may still be applied; nothing retries it on its own. Pass deadline: with the write: once the deadline plus the largest difference between the nodes' clocks has passed, the write can no longer be applied. A read with consistency: :linearizable then returns a state that the write can no longer change. The row alone shows whether the write was applied only if no other writer could have changed it. To tell in every case, write a receipt in the same batch: a row under a key unique to the request, in the same partition, that nothing else deletes. After the deadline, the write was applied if and only if a linearizable read finds the receipt.
  • Scans. Each page reads one consistent view of the durable rows. Pages are separate reads, so rows written between pages can appear in later pages if their keys come later. Passing cursor: nil starts at the beginning of the selected range; a returned cursor: nil means iteration is complete. A binary cursor identifies the last key returned and remains valid after that row is deleted. Supplied :prefix, :gte, and :lt values must be binaries. Scan rows omit versions by default; with_versions: true includes them. There are no scans or writes across partitions.

Atomic batches

Slap.KV.batch/3 applies operations on distinct keys in one partition atomically:

{:ok, version} = Slap.KV.batch("user:42", [
  {:put, "profile", "Ada", :absent},
  {:put, "theme", "dark"},
  {:delete, "draft"},
  {:check, "disabled", :absent}
])

The rules for operations:

  • A put or delete may end with a version condition: an integer version, or, for a put, :absent.
  • A check must have a condition: an integer version or :absent.
  • Each key may appear only once in a batch, so a key cannot be both checked and changed. A condition of nil is invalid.
  • Conditions are checked in list order against the rows as they were before the batch. The first one that fails returns {:error, {:conflict, key, current_version_or_nil}}, and nothing is written.

A batch that puts or deletes rows returns {:ok, version}, and every row it puts gets that version. A batch of only checks returns {:ok, nil}; like a conflict, it replies only after a write of its own confirms that its node still owns the shard. Options are :deadline, :timeout and :cluster, as for single writes. A batch is not applied at or after its deadline. A timeout does not cancel the batch: after :timeout or :unavailable, resolve its outcome as for single writes (see Uncertain writes) before retrying.

Limits on pending work

Each shard limits how much work can wait for durability at once. A write, batch or linearizable get counts against the limits from the moment its partition writer accepts it until its result is durable and it has replied. A request that would exceed them is rejected before any row is read or written. Set the limits as positive integers in the cluster's child_options:

child_options: [
  partition_writers: 16,
  max_inflight_bytes_per_writer: 64 * 1024 * 1024,
  max_inflight_bytes_per_shard: 256 * 1024 * 1024,
  max_inflight_requests_per_writer: 1_024,
  max_batch_ops: 1_000,
  max_batch_bytes: 8 * 1024 * 1024
]

These are the defaults. A request's size in bytes is the sum, over its operations, of the encoded key (two length bytes, the partition and the key) and, for a put, the value. A batch counts as one request. A linearizable get counts as one request of zero bytes. Durable gets and scans are not limited.

  • A request larger than either byte limit returns {:error, {:bad_request, :request_too_large}}, even when nothing else is pending.
  • A batch with more than max_batch_ops operations or max_batch_bytes bytes returns {:error, {:bad_request, :batch_too_large}}.
  • A request that would fit once pending work finishes returns {:error, :overloaded}.

In all three cases the request was not applied.

A caller's timeout or exit does not free the work's capacity: the work is still pending. If a partition writer crashes, its pending writes count until they are durable or the shard's database closes. If it crashes while submitting a write to SlateDB, so that whether the write was submitted is unknown, the shard stops, and it accepts requests again only after its database has closed.

These limits bound pending request bytes and counts. They do not bound process mailboxes, HTTP decoding, distribution queues, response bodies, or total BEAM or native memory. Slap.KV.Telemetry reports rejections and pending work.

How it works

  • Placement. A partition is on shard xxh64(partition) mod N (Slap.KV.Cluster.shard_for/1). Operations run on the shard's owner through Slap.Cluster.call/4.
  • Writers. Each shard's database has one SlateDB writer: the handle its owner opened. slap_cluster places the owner on one node, and SlateDB fences any other. On that node, each partition has one partition writer, the only process that writes the partition's rows. Each shard runs partition_writers partition writers (a cluster child_options key, 16 by default), and all of them write through the shard's one handle. A partition belongs to the partition writer :erlang.phash2(partition, partition_writers). A shard reads the setting when it starts its partition writers, and keeps that count until it starts them again.
  • Conditions and pending writes. A partition writer checks a condition with a normal read, which sees its own writes even before they are durable. It writes without waiting and replies once the write is durable, so several writes can be in flight at once. Replies go out in request order, behind the writes in flight.
  • Conflicts are confirmed. Under a placement without leases, a node can lose a shard after a pause without knowing it. Its partition writers then judge conditions from old state. So before reporting a conflict, a partition writer writes a reserved key and waits for that write to be durable. On a fenced handle the write fails, and the request gets :unavailable instead of a conflict that may be wrong. Successful writes need no confirmation, because they fail on a fenced handle anyway. Reads on such a node may be stale until it finds that it has been fenced, as with streams.
  • Writer recovery. A write error fails every request in flight with :unavailable and stops the partition writer. Its supervisor starts a new one, which first waits until every write made through the shard's database is durable. So the new writer never checks a condition against state that could still be lost.

HTTP

Slap.KV.HTTP.Router serves KV over HTTP, under /v1/kv by default. To try it, download and run the standalone server script from slap in another terminal:

curl -fsSLo server.exs https://raw.githubusercontent.com/mjrusso/slap/main/slap/examples/server.exs
elixir server.exs

It keeps its data in memory, which is lost when the server stops. Then:

curl -X PUT --data-binary 'hello' -H 'If-None-Match: *' -i http://localhost:4437/v1/kv/p/greeting
curl -i http://localhost:4437/v1/kv/p/greeting
curl 'http://localhost:4437/v1/kv/p?prefix=gr&limit=10'

The PUT returns 204 and an ETag containing the new version; the GET returns 200 with the value and the same ETag. To replace or delete the row, send that value in If-Match. A stale version returns 412 with the current ETag.

Partitions and keys are percent-encoded path segments. In scan responses, keys, values, and the cursor are unpadded base64url. The router does not authenticate requests.

Size errors return 413 without Retry-After. A 503 response with body overloaded: request was not applied is a definite rejection and includes Retry-After: 1. Other 503 responses and transport failures leave the write's outcome unknown.

For an atomic batch, send JSON with unpadded base64url keys and values:

curl -i -X POST http://localhost:4437/v1/kv/p \
  -H 'Content-Type: application/json' \
  --data '{"ops":[{"op":"put","key":"YQ","value":"MQ","if_absent":true},{"op":"delete","key":"Yg"}]}'

Operations are put (key and value), delete (key), or check (key and a required condition). Conditions are if_version: N or if_absent: true; if_absent is allowed only for put and check. A batch is rejected if it has unknown fields, an operation has both conditions, or the request has an If-Match or If-None-Match header. The only optional top-level field is integer deadline in Unix milliseconds.

Mutation success is 200 with {"version": N} and ETag "N". Check-only success is 200 with {"version": null} and no ETag. A conflict is 412 with {"key": "base64url", "version": N} (null when missing). Passed batch deadlines return 400 with deadline_exceeded.

The router's max_batch_body bounds raw JSON (16 MiB by default), separately from owner-enforced decoded limits. max_value also bounds each decoded put value (8 MiB by default).

Storage

A row's SlateDB key is the partition's length (16 bits), the partition, then the key, so a partition is one contiguous key range and scans never cross into another partition. Keys may be up to SlateDB's limit of 65,535 bytes once encoded. Deletes are SlateDB deletes (tombstones, removed by compaction).

Limits

  • A conditional write of a row that is not in memory waits for a read from the store, and its partition writer's other writes wait behind it. Unconditional writes do not read. Partition writers are independent, so this delays only the partitions that share that writer; see Benchmarks.
  • No TTLs or cross-partition operations.

Benchmarks

Follow the repository's development setup, then run from slap_kv/:

mix run bench/partition_writers.exs --store memory

bench/partition_writers.exs measures throughput, latency, and writer mailbox length with unconditional, warm conditional, and cold conditional writes. A cold conditional write waits for a store read and blocks its partition writer; increasing partition_writers spreads that delay across partitions. Measure on the intended store before changing the default of 16.

bench/durability_admission.exs runs 64 bounded clients by default, mixing 64 KiB puts, tiny puts, failed conditions and linearizable gets. It reports elapsed time, request outcomes, reserved bytes/counts, mailbox samples, BEAM memory and SlateDB's native buffer metrics as JSON. Native buffer gauges cover only the memory SlateDB reports; they are not total native allocations. Use an externally delayed object store to measure slow durability. The report also includes the result of a final write after the clients finish.

mix run bench/durability_admission.exs --store s3://bucket/benchmark --flush-interval 10ms

Tests

Run from slap_kv/ after the development setup above:

mix test          # SLAP_KV_PROP_RUNS=1000 for more model sequences

The model test covers atomic batches, conditional writes, scans, writer replacement, and cluster restarts. test/durability_test.exs checks durable replies, pending writes, and fencing. Set SLAP_KV_PROP_RUNS to run more model sequences.

License

Slap is released under the terms of the Apache License 2.0.

Copyright (c) 2026, Michael Russo.