Atomic Bucket

Fast single node rate limiter implementing Token Bucket algorithm. The goal is to provide dependable solution that JustWorks™ with a focus on performance, correctness and ease of use. Bucket data is stored using :atomics module. Bucket references are stored in ETS and optionally cached as persistent terms.

Features

lock-free and race-free with compare-and-swap operations
BlazingFast™ performance, see benchmarks section. Req/s go brrrrrr
monotonic timer for correct calculations
millisecond tick supporting wider range of parameters and preventing request starvation
automatic calculation of bucket parameters based on average rate and burst size
handy timeouts for retries
support for token refunds and variable cost requests
compile-time validation of arguments when possible

Installation

Add it to your list of dependencies in mix.exs and run mix deps.get:

def deps do
  [
    {:atomic_bucket, "~> 0.2"}
  ]
end

You must start AtomicBucket server for each bucket table you want to use - without it the library will not work. Add to your application:

children = [.., AtomicBucket, ..]

This will once per hour clean buckets that haven't had requests in the last 24 hours. See start_link/1 for info about available options.

Usage

For simple cases where requests have fixed cost use request/5 macro with desired average rate and burst parameters. When possible, call the macro with literal arguments for better performance and compile-time validation. Module attributes are fine too.

require AtomicBucket

# Averate rate: 10 reqs/s with 3 burst requests. 
case AtomicBucket.request(:mybucket, 1, 10, 3) do
  {:allow, count, _ref} ->
    # Request is allowed. May immediately attempt to make additional
    # <count> calls.
  {:deny, timeout, _ref} ->
    # Request is denied. The bucket may have enough tokens in <timeout>
    # milliseconds.
end

Bucket id can be any term.

AtomicBucket.request({:client, ip_addr}, 1, 10, 3)

Cache bucket reference in :persistent_term for better performance. Good fit for buckets with low churn. Best for fixed buckets like per-user-id rate limits. See :persistent_term docs for more info on the tradeoffs.

AtomicBucket.request(:mybucket, 1, 10, 3, persistent: true)

Reuse bucket references in long running processes for top performance.

{:allow, _requests, bucket_ref} = AtomicBucket.request(:mybucket, 1, 10, 3)
AtomicBucket.request(:mybucket, 1, 10, 3, ref: bucket_ref)

Use raw_request/5 macro to implement advanced features such as token refunds or variable cost. It supports same options as request/5

# This would be 10 req/s with 2 burst requests in fixed cost scenario
{:allow, remaining_tokens, ref} = AtomicBucket.raw_request(:mybucket, 200, 1, 100)
# But the next request may have a different cost
AtomicBucket.raw_request(:mybucket, 200, 1, 150)
# Token "refund" is always allowed
AtomicBucket.raw_request(:mybucket, 200, 1, -100)

To implement different retention policies start multiple servers and use the table option of request/5 and raw_request/5. Bucket ids are table-scoped and don't have to be globally unique.

Caveats

The library is optimized for common cases where rate limiters are used. Extremely slow/fast rates and/or huge bursts may exceed the bucket storage limits (64 bits). In practice, most people wouldn't need these extreme parameters.

The library makes no effort to ensure that bucket parameters remain stable across calls: the parameters are not stored at all! Using same bucket with different parameters will result in silent bugs. This also applies to mixing request/5 and raw_request/5 - it must be avoided.

Benchmarks

The library provides a benchmark measuring series of 1000 rate limit checks (see bench/benchmark.exs). It serves as an illustration of available options. In general, such benchmarks should be taken with a grain of salt:

they often use artificial conditions different from real life, thus devaluing the results
above certain point a rate limiter is "fast enough" for many purposes. For BEAM this probably means "ETS or faster". Any atomics-based library should be much faster than ETS and is likely the fastest you can get.
comparing results between differently implemented or completely different algorithms is neither 100% valid nor very useful. It's very easy to make a pointless benchmark measuring wrong things, simply because it's hard to make different solutions perform same work achieving same results.
Tradeoffs are important. It's better to look at the overall picture, rather than raw performance.

$ mix run bench/benchmark.exs
Operating System: Linux
CPU Information: 13th Gen Intel(R) Core(TM) i7-13700K
Number of Available Cores: 24
Available memory: 62.52 GB
Elixir 1.20.0
Erlang 29.0.1
JIT enabled: true

Benchmark suite executing with the following configuration:
warmup: 2 s
time: 5 s
memory time: 0 ns
reduction time: 0 ns
parallel: 1
inputs: none specified
Estimated total run time: 42 s
Excluding outliers: false

Benchmarking raw_request (literals, default opts) ...
Benchmarking raw_request (non-literals, default opts) ...
Benchmarking request (literals, default opts) ...
Benchmarking request (literals, persistent) ...
Benchmarking request (literals, reusing ref) ...
Benchmarking request (non-literals, default opts) ...
Calculating statistics...
Formatting results...

Name                                               ips        average  deviation         median         99th %
request (literals, reusing ref)                11.81 K       84.65 μs    ±13.79%       82.21 μs      135.81 μs
request (literals, persistent)                  8.42 K      118.71 μs    ±11.83%      115.87 μs      193.22 μs
raw_request (literals, default opts)            5.31 K      188.18 μs    ±12.08%      183.14 μs      308.21 μs
request (literals, default opts)                5.29 K      188.90 μs    ±11.78%      184.10 μs      306.53 μs
raw_request (non-literals, default opts)        5.00 K      200.19 μs    ±11.56%      195.08 μs      321.36 μs
request (non-literals, default opts)            4.27 K      234.02 μs    ±10.90%      227.43 μs      353.50 μs

Comparison: 
request (literals, reusing ref)                11.81 K
request (literals, persistent)                  8.42 K - 1.40x slower +34.06 μs
raw_request (literals, default opts)            5.31 K - 2.22x slower +103.53 μs
request (literals, default opts)                5.29 K - 2.23x slower +104.26 μs
raw_request (non-literals, default opts)        5.00 K - 2.37x slower +115.55 μs
request (non-literals, default opts)            4.27 K - 2.76x slower +149.37 μs

Extended statistics: 

Name                                             minimum        maximum    sample size                     mode
request (literals, reusing ref)                 77.47 μs      369.01 μs        58.94 K79.16 μs, 79.58 μs, 79.75
request (literals, persistent)                 110.33 μs      378.21 μs        42.05 K     112.17 μs, 112.11 μs
raw_request (literals, default opts)           164.90 μs      464.82 μs        26.54 K                187.78 μs
request (literals, default opts)               164.81 μs      418.69 μs        26.44 K     186.95 μs, 187.26 μs
raw_request (non-literals, default opts)       176.23 μs      470.45 μs        24.95 K198.13 μs, 189.40 μs, 198
request (non-literals, default opts)           206.55 μs      434.20 μs        21.34 K227.04 μs, 224.74 μs, 227

License

Copyright 2026 Andrey Tretyakov
The source code of the project is released under Apache License 2.0. Check LICENSE file for more information.