slap_cluster

slap_cluster runs a fixed number of SlateDB databases, called shards, across the nodes of an Elixir cluster. The shards share one store, such as a prefix in an S3 bucket. slap_cluster assigns each shard to one node, opens the shard's database there, and routes calls for the shard to that node. Depending on the placement strategy, a failed node's shards move to other nodes. SlateDB allows one writer per database: if two nodes open the same shard, only the newer one can write. slap_cluster also reports when a shard's writes become durable. It opens the databases with slap_slatedb.

An object store must support conditional writes, as S3 does. The cluster checks this when it starts, and does not start if the store fails the check.

slap_cluster does not decide what the shards store. Your application provides the processes that run with each shard on its node, called shard children, and they implement the service's reads, writes, and replies. For one database on one node, use slap_slatedb directly. Use slap_cluster directly if your data does not fit an existing service such as slap_streams or slap_kv (which are themselves built on slap_cluster).

About Slap {: .info}

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

Standalone server:

slap: an HTTP server for evaluating and benchmarking slap_streams and slap_kv

Installation

Add slap_cluster to the dependencies in your application's mix.exs:

defp deps do
  [
    {:slap_cluster, "~> 0.2.0"}
  ]
end

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

Example

In an application that depends on slap_cluster, define a cluster module:

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

Configure it in config/config.exs. This local configuration sketch assumes the application provides MyApp.ShardChildren.child_specs/1 to start its processes for each shard:

config :my_app, MyApp.Cluster,
  store: {:local, "/tmp/my-app-cluster"},
  shards: 8,
  shard_children: {MyApp.ShardChildren, :child_specs, []}

Add MyApp.Cluster to the application's supervision tree after defining its shard children.

In a shard child, ctx is its context, ops is a SlateDB write batch, and ref is a token chosen by the child to identify the notification. The child can request a durability notification for a write:

{:ok, seq} = Slap.SlateDB.write(ctx.db, ops)
Slap.Cluster.notify_when_durable(ctx, seq, {:ack, ref})

The child later receives {:slap_cluster_durable, {:ack, ref}}. A caller can route an operation to the owner of a key's shard. Here MyApp.Streams.append/2 is an application function, and path and body are its arguments:

shard = MyApp.Cluster.shard_for("some/key")
{:ok, result} = MyApp.Cluster.call(shard, {MyApp.Streams, :append, [path, body]})

The outer {:ok, result} means routing succeeded. result is the application function's return value, including any {:error, reason} it returns.

Slap.Cluster's docs cover the options, shard children, durability and telemetry.

How it works

The cluster probes the store's create-if-absent behavior before opening shards. ObjectLease also requires If-Match support. Each shard has one database handle and one durability subscription. Application shard children stop before the handle closes. A fenced or crashed database stops its shard; the placement strategy decides where to reopen it. Reopening without that decision could fence a new owner. Children that need the stop reason use shard_status/1; their supervisor exit reason is :shutdown even when the shard was fenced.

When a child calls notify_when_durable/3, the shard compares the write's sequence number with SlateDB's current durable sequence number, not with the last durability message it handled. A write that is already durable is reported at once, even if that message is still in the shard's mailbox. shard_for/1 uses xxh64(key) mod shards; changing the shard count or hash moves existing keys.

Several nodes

To run a cluster on several nodes, use a shared object store and a placement strategy. For an S3-compatible store such as RustFS, aws_endpoint sets its endpoint; the store must support conditional writes:

config :my_app, MyApp.Cluster,
  store:
    {:url, "s3://bucket/prefix",
     aws_endpoint: "http://rustfs:9000", aws_allow_http: true},
  shards: 64,
  settings: %{flush_interval: "10ms"},
  cache: [capacity_bytes: 2 * 1024 * 1024 * 1024],
  shard_children: {MyApp.ShardChildren, :child_specs, []},
  strategy: {Slap.Cluster.Strategy.ObjectLease, lease_ttl: 15_000}

Nodes connect through distributed Erlang. ObjectLease stores conditional leases and node heartbeats in the bucket. An owner renews its leases and stops a shard if renewal fails past its local deadline, before another node can claim it. A paused owner that misses the deadline is fenced by SlateDB when a new owner opens the database. A clean stop closes shards before freeing their leases. Self-fencing runs separately from the renewal loop so a blocked store call cannot delay it.

call/4 routes to the owner over :erpc. If the reply is :not_owner or :unassigned, or the owner cannot be reached before the call is sent, call/4 refreshes the placement and retries. A call that was already sent is not retried, because it may have run; the caller receives {:error, {:erpc, :noconnection}}. Lease data does not create node-name atoms, so peers must be discovered or configured before routing to them.

Placement strategies

Local runs all shards on one node. The node must restart to recover from a failure.
ObjectLease uses an object store with If-Match (S3, RustFS, GCS or Azure) and roughly synchronised clocks. Failover takes about one lease TTL. A node cut off from the bucket stops its shards before its leases expire.
Distributed uses distributed Erlang without leases. A dead node's shards move after :settle (2 s); an unresponsive node also requires net_ticktime. A partition can let both sides open a shard. SlateDB fences one writer, but that side may serve stale reads until it finds that it has been fenced.
Static uses a fixed list of nodes. A failed node must return before its shards are available again.

ObjectLease requires clocks within :max_clock_skew_ms (1 s by default); the local file store cannot provide If-Match. Lower net_ticktime with Distributed if failure detection must be faster.

Tests

mix test                                                     # local directories
SLAP_TEST_S3_ENDPOINT=http://127.0.0.1:9000 mix test      # also 64 shards on RustFS

The tests exercise durability, fencing, placement and routing with a counter server on each shard. Multi-node tests use separate OS processes and a shared store; SLAP_TEST_S3_ENDPOINT enables the RustFS cases. See test/support/failover.ex for the fault test harness.

License

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