StableInterfaces

Consistently routable interface instances in Go without the need for storage or coordination. Inspired by the use case for Cloudflare Durable Objects.

StableInterfaces

Summary

Stable Interfaces (SIs) provide scaffolding for building consistently-routable instances of Go interfaces across one or more machines. They guarantee a single instance of an interface is running (if you deploy correctly, sacrificing availability for consistency), and all do so without the need for storage, consensus, or coordination

Stable Interfaces (SIs) vs DurableObjects (DOs)

DOs are fantastic, and cover >90% of use cases. There are some cases where they simply don't work, and this is where SIs excel:

Go interfaces, extreme flexibility
Max resources per instance is the host's resources
Get a full OS underneath
Run anywhere (in your api, airgapped, separate service)
Connect to existing DB for persistence (all your data in one spot)
Better alarms
Use any networking: TCP, UDP, GRPC, HTTP/(1,2,3)
Unlimited concurrent connections: DO's have a hard 32k limit (I've hit it in production)
Cheaper per-unit cost of runtime and requests at scale

If you can use DOs, you probably should. I've only found a few bugs in them 🙃.

Hosts

Hosts must be named sequentially (e.g. host-1, host-2, etc.). This is very simple to do in system like Kuberenetes with StatefulSets (pro-tip: use podManagementPolicy: Parllel when possible), and aren't too terrible using metadata on systems like fly.io. This is because shards are deterministically assigned to hosts in sequential order (e.g. host-1 gets shard 1, host-2 gets shard 2, and so on in a circle).

Connection Handling

Synchronous `Send()`

The Send() does not wait for a response, it launches the OnRecv() handler in a goroutine to release closure locks.

If you want to wait for the result of Send() (perhaps you want to verify that either the interface or client handled the message), you can simply pass an included chan error that you listen for before exiting.

Alarms

You can optionally enable alarms by passing WithAlarm() to NewInterfaceManager:

NewInterfaceManager(host, "host-{0..1}", 1024, func(internalID string) StableInterface {
    // ...
}, WithAlarm(AlarmManager))

The AlarmManager is an interface that implements a few functions for persisting and recovering from restarts. It is expected that you use a consistent key-value store. Some examples of systems in which you can do this:

Postgres/CockroachDB, MySQL
Cassandra/ScyllaDB with LWTs or (n/2)+1 consistency level
Redis sorted sets with AOF persistence
FoundationDB, etcd, and other consistent KV stores

Alarms are managed per-shard, and mostly write to the database (creation and deletion). Reads only occur when a node restarts and needs to read in any previously created alarms.

Unlike DurableObjects, alarms are a bit more capable in StableInterfaces. Each alarm has an ID and Meta available to it. This means you can make multiple alarms at the same time, which will fire off in (time, ID) order. Meta is a map[string]any (must be serializable), so you can attach metadata to your alarm to know what it's for. You can also list, update, and cancel alarms. This is a simple pattern for durable background processing.

By default, alarms will be tested for every 150 milliseconds. You can override this with the WithAlarmCheckInterval() option. Alarms are handled sequentially, one at a time. This interval is only used between checks of no active alarms. If an alarm fires, the handler is launched in a goroutine and the alarm check immediately runs.

Alarms also have configurable backoff as well, see interface_manager_options.go.

It's important to note that when started, every shard will query the AlarmManager for the latest alarms. You may want to introduce some form of rate limiting if you are unable to handle the burst of query activity. See #3 (comment) for more. It currently loads all pending alarms in memory, so make sure you don't have billions of alarms. Each alarm is lightweight (just a struct in a tree), so the only thing you have to worry about is memory size.

Other tips

Persistence of events

A primary use case is some form of stateful instance/machine. In fact, one of the major reasons I made this was because DOs were getting really expensive at >1B req/month, and Temporal (as nice as it is) has far too much overhead for the majority of our use cases.

Many of our use cases are some variation of "When event X happens, wait Y for event Z, and do something depending on whether that times out".

However, you still want to store some persistence for instances such as recovery from restarts and recovery. You most likely also still want to guarantee read-after-write, and durability with handlers like OnRequest().

ClickHouse as an event store

The most efficient mechanism we've found for this is as follows:

When you receive an event you want to persist, write it to a batch-writer for ClickHouse (order like instance_id,event_id, event_id being sorted)
When that writes, process the event and respond to the OnRequest() handler. Consider how you might want to queue the events in-memory so that if multiple events come in and write to the same batch, you process them in the order they arrived to the instance
Keep some flag in memory to indicate whether you just booted (use a mutex, so you can hold new incoming events), and back-fill in the events from ClickHouse by streaming the rows in. Make sure you have some flag indicating you are back-filling, so you know not to fire things such as emails while doing this.

You can make modifications such as not waiting for writes, async batch writing ClickHouse-side (optionally waiting), not queueing events based on time in memory, and more. Make sure you use the native protocol and stream rows for reads.

To support multiple hosts, see the Hash-based section below

Snapshot-style

Another method is to perform snapshots if your instance. This could be done at some max interval (e.g. every 10 seconds if state has changed), or far more frequently (such as every write).

Reconstructing from the original event log is preferable (i.e. ClickHouse event store), however for some other cases this may be more appropriate

Hybrid

Multi-node Coordination

Hash-based

An instance ID maps to a node determined by a hashing function.

You can have crude verisons of this (k8s stateful sets), or more complex versions (supporting hybird IDs so you can have part ID map to a cluster, and part be the instance ID, then can cache lookups for cluster ID in memory).

To support recovery (persistence), you can modify the ClickHouse method above to add a materialized view that performs the same hashing function as your routing logic, resulting in a table like hash_token,instance_id,event_id. Then, when a node that has a token assigned to it recovers (e.g. the *-1 statefulset pod boots in k8s), it can easily pull all of the data for that hash token.

Depending on how often you fail-over, it might be better to drop the event here, and only store the hash_token,instance_id mapping (omit the event data), then can resolve the event data by pulling the event feed by instance_id later.

This makes it relatively trivial to increase and decrease the number of hash tokens, because you can just rebuild the materialized view in the background, switch once built, fail-over nodes, then when everything is recovered you can delete the old materialized view.

There are many other methods, this is just a particularly lightweight and easy method to implement at scale if it satisfies requirements.

danthegoodman1/StableInterfaces

ErrorLooks like something went wrong!

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Languages

Error
Looks like something went wrong!

Packages