millione / moka

A high performance concurrent caching library for Rust

Moka

Moka is a fast, concurrent cache library for Rust. Moka is inspired by the Caffeine library for Java.

Moka provides cache implementations on top of hash maps. They support full concurrency of retrievals and a high expected concurrency for updates.

All caches perform a best-effort bounding of a hash map using an entry replacement algorithm to determine which entries to evict when the capacity is exceeded.

Features

Moka provides a rich and flexible feature set while maintaining high hit ratio and a high level of concurrency for concurrent access.

• Thread-safe, highly concurrent in-memory cache implementations:
  • Synchronous caches that can be shared across OS threads.
  • An asynchronous (futures aware) cache that can be accessed inside and outside of asynchronous contexts.
• A cache can be bounded by one of the followings:
  • The maximum number of entries.
  • The total weighted size of entries. (Size aware eviction)
• Maintains near optimal hit ratio by using an entry replacement algorithms inspired by Caffeine:
  • Admission to a cache is controlled by the Least Frequently Used (LFU) policy.
  • Eviction from a cache is controlled by the Least Recently Used (LRU) policy.
  • More details and some benchmark results are available here.
• Supports expiration policies:
  • Time to live
  • Time to idle
• Supports eviction listener, a callback function that will be called when an entry is removed from the cache.

Choosing the right cache for your use case

No cache implementation is perfect for every use cases. Moka is a complex software and can be overkill for your use case. Sometimes simpler caches like Mini Moka or Quick Cache might be a better fit.

The following table shows the trade-offs between the different cache implementations:

Feature	Moka v0.10	Mini Moka v0.10	Quick Cache v0.2
Thread-safe, sync cache	✅	✅	✅
Thread-safe, async cache	✅	❌	❌
Non-concurrent cache	❌	✅	✅
Bounded by the maximum number of entries	✅	✅	✅
Bounded by the total weighted size of entries	✅	✅	✅
Near optimal hit ratio	✅ TinyLFU	✅ TinyLFU	✅ CLOCK-Pro
Expiration policies	✅	✅	❌
Eviction listener	✅	❌	❌
Per-key, atomic insertion	✅ get_with family methods	❌	❌
Lock-free, concurrent iterator	✅	❌	❌
Lock-per-shard, concurrent iterator	❌	✅	❌

Performance	Moka v0.10	Mini Moka v0.10	Quick Cache v0.2
Small overhead compared to a concurrent hash table	❌	❌	✅
Does not use background threads	❌ Will be removed from v0.11	✅	✅
Small dependency tree	❌	✅	✅

Moka in Production

Moka is powering production services as well as embedded Linux devices like home routers. Here are some highlights:

• crates.io: The official crate registry has been using Moka in its API service to reduce the loads on PostgreSQL. Moka is maintaining cache hit rates of ~85% for the high-traffic download endpoint. (Moka used: Nov 2021 — present)
• aliyundrive-webdav: This WebDAV gateway for a cloud drive may have been deployed in hundreds of home Wi-Fi routers, including inexpensive models with 32-bit MIPS or ARMv5TE-based SoCs. Moka is used to cache the metadata of remote files. (Moka used: Aug 2021 — present)

Change Log

• CHANGELOG.md

The unsync::Cache and dash::Cache have been moved to a separate crate called Mini Moka:

• moka::unsync::Cache → mini_moka::unsync::Cache
• moka::dash::Cache → mini_moka::sync::Cache

Table of Contents

• Features
  • Choosing the right cache for your use case
• Moka in Production
• Change Log
• Usage
• Examples (Part 1)
  • Synchronous Cache
  • Asynchronous Cache
• Avoiding to clone the value at get
• Examples (Part 2)
  • Size Aware Eviction
  • Expiration Policies
• Hashing Algorithm
• Minimum Supported Rust Versions
• Troubleshooting
  • Integer Overflow in Quanta Crate on Some x86_64 Machines
  • Compile Errors on Some 32-bit Platforms
• Developing Moka
• Road Map
• About the Name
• Credits
• License

Usage

Add this to your Cargo.toml:

[dependencies]
moka = "0.10"

To use the asynchronous cache, enable a crate feature called "future".

[dependencies]
moka = { version = "0.10", features = ["future"] }

Example: Synchronous Cache

The thread-safe, synchronous caches are defined in the sync module.

Cache entries are manually added using insert or get_with method, and are stored in the cache until either evicted or manually invalidated.

Here's an example of reading and updating a cache by using multiple threads:

// Use the synchronous cache.
use moka::sync::Cache;

use std::thread;

fn value(n: usize) -> String {
    format!("value {}", n)
}

fn main() {
    const NUM_THREADS: usize = 16;
    const NUM_KEYS_PER_THREAD: usize = 64;

    // Create a cache that can store up to 10,000 entries.
    let cache = Cache::new(10_000);

    // Spawn threads and read and update the cache simultaneously.
    let threads: Vec<_> = (0..NUM_THREADS)
        .map(|i| {
            // To share the same cache across the threads, clone it.
            // This is a cheap operation.
            let my_cache = cache.clone();
            let start = i * NUM_KEYS_PER_THREAD;
            let end = (i + 1) * NUM_KEYS_PER_THREAD;

            thread::spawn(move || {
                // Insert 64 entries. (NUM_KEYS_PER_THREAD = 64)
                for key in start..end {
                    my_cache.insert(key, value(key));
                    // get() returns Option<String>, a clone of the stored value.
                    assert_eq!(my_cache.get(&key), Some(value(key)));
                }

                // Invalidate every 4 element of the inserted entries.
                for key in (start..end).step_by(4) {
                    my_cache.invalidate(&key);
                }
            })
        })
        .collect();

    // Wait for all threads to complete.
    threads.into_iter().for_each(|t| t.join().expect("Failed"));

    // Verify the result.
    for key in 0..(NUM_THREADS * NUM_KEYS_PER_THREAD) {
        if key % 4 == 0 {
            assert_eq!(cache.get(&key), None);
        } else {
            assert_eq!(cache.get(&key), Some(value(key)));
        }
    }
}

If you want to atomically initialize and insert a value when the key is not present, you might want to check the document for other insertion methods get_with and try_get_with.

Example: Asynchronous Cache

The asynchronous (futures aware) cache is defined in the future module. It works with asynchronous runtime such as Tokio, async-std or actix-rt. To use the asynchronous cache, enable a crate feature called "future".

Cache entries are manually added using an insert method, and are stored in the cache until either evicted or manually invalidated:

• Inside an async context (async fn or async block), use insert or invalidate method for updating the cache and await them.
• Outside any async context, use blocking method to access blocking version of insert or invalidate methods.

Here is a similar program to the previous example, but using asynchronous cache with Tokio runtime:

// Cargo.toml
//
// [dependencies]
// moka = { version = "0.10", features = ["future"] }
// tokio = { version = "1", features = ["rt-multi-thread", "macros" ] }
// futures-util = "0.3"

// Use the asynchronous cache.
use moka::future::Cache;

#[tokio::main]
async fn main() {
    const NUM_TASKS: usize = 16;
    const NUM_KEYS_PER_TASK: usize = 64;

    fn value(n: usize) -> String {
        format!("value {}", n)
    }

    // Create a cache that can store up to 10,000 entries.
    let cache = Cache::new(10_000);

    // Spawn async tasks and write to and read from the cache.
    let tasks: Vec<_> = (0..NUM_TASKS)
        .map(|i| {
            // To share the same cache across the async tasks, clone it.
            // This is a cheap operation.
            let my_cache = cache.clone();
            let start = i * NUM_KEYS_PER_TASK;
            let end = (i + 1) * NUM_KEYS_PER_TASK;

            tokio::spawn(async move {
                // Insert 64 entries. (NUM_KEYS_PER_TASK = 64)
                for key in start..end {
                    // insert() is an async method, so await it.
                    my_cache.insert(key, value(key)).await;
                    // get() returns Option<String>, a clone of the stored value.
                    assert_eq!(my_cache.get(&key), Some(value(key)));
                }

                // Invalidate every 4 element of the inserted entries.
                for key in (start..end).step_by(4) {
                    // invalidate() is an async method, so await it.
                    my_cache.invalidate(&key).await;
                }
            })
        })
        .collect();

    // Wait for all tasks to complete.
    futures_util::future::join_all(tasks).await;

    // Verify the result.
    for key in 0..(NUM_TASKS * NUM_KEYS_PER_TASK) {
        if key % 4 == 0 {
            assert_eq!(cache.get(&key), None);
        } else {
            assert_eq!(cache.get(&key), Some(value(key)));
        }
    }
}

If you want to atomically initialize and insert a value when the key is not present, you might want to check the document for other insertion methods get_with and try_get_with.

Avoiding to clone the value at get

For the concurrent caches (sync and future caches), the return type of get method is Option<V> instead of Option<&V>, where V is the value type. Every time get is called for an existing key, it creates a clone of the stored value V and returns it. This is because the Cache allows concurrent updates from threads so a value stored in the cache can be dropped or replaced at any time by any other thread. get cannot return a reference &V as it is impossible to guarantee the value outlives the reference.

If you want to store values that will be expensive to clone, wrap them by std::sync::Arc before storing in a cache. Arc is a thread-safe reference-counted pointer and its clone() method is cheap.

use std::sync::Arc;

let key = ...
let large_value = vec![0u8; 2 * 1024 * 1024]; // 2 MiB

// When insert, wrap the large_value by Arc.
cache.insert(key.clone(), Arc::new(large_value));

// get() will call Arc::clone() on the stored value, which is cheap.
cache.get(&key);

Example: Size Aware Eviction

If different cache entries have different "weights" — e.g. each entry has different memory footprints — you can specify a weigher closure at the cache creation time. The closure should return a weighted size (relative size) of an entry in u32, and the cache will evict entries when the total weighted size exceeds its max_capacity.

use std::convert::TryInto;
use moka::sync::Cache;

fn main() {
    let cache = Cache::builder()
        // A weigher closure takes &K and &V and returns a u32 representing the
        // relative size of the entry. Here, we use the byte length of the value
        // String as the size.
        .weigher(|_key, value: &String| -> u32 {
            value.len().try_into().unwrap_or(u32::MAX)
        })
        // This cache will hold up to 32MiB of values.
        .max_capacity(32 * 1024 * 1024)
        .build();
    cache.insert(0, "zero".to_string());
}

Note that weighted sizes are not used when making eviction selections.

Example: Expiration Policies

Moka supports the following expiration policies:

• Time to live: A cached entry will be expired after the specified duration past from insert.
• Time to idle: A cached entry will be expired after the specified duration past from get or insert.

To set them, use the CacheBuilder.

use moka::sync::Cache;
use std::time::Duration;

fn main() {
    let cache = Cache::builder()
        // Time to live (TTL): 30 minutes
        .time_to_live(Duration::from_secs(30 * 60))
        // Time to idle (TTI):  5 minutes
        .time_to_idle(Duration::from_secs( 5 * 60))
        // Create the cache.
        .build();

    // This entry will expire after 5 minutes (TTI) if there is no get().
    cache.insert(0, "zero");

    // This get() will extend the entry life for another 5 minutes.
    cache.get(&0);

    // Even though we keep calling get(), the entry will expire
    // after 30 minutes (TTL) from the insert().
}

A note on expiration policies

The cache builders will panic if configured with either time_to_live or time to idle longer than 1000 years. This is done to protect against overflow when computing key expiration.

Hashing Algorithm

By default, a cache uses a hashing algorithm selected to provide resistance against HashDoS attacks.

The default hashing algorithm is the one used by std::collections::HashMap, which is currently SipHash 1-3, though this is subject to change at any point in the future.

While its performance is very competitive for medium sized keys, other hashing algorithms will outperform it for small keys such as integers as well as large keys such as long strings. However those algorithms will typically not protect against attacks such as HashDoS.

The hashing algorithm can be replaced on a per-Cache basis using the build_with_hasher method of the CacheBuilder. Many alternative algorithms are available on crates.io, such as the AHash crate.

Minimum Supported Rust Versions

Moka's minimum supported Rust versions (MSRV) are the followings:

Feature	MSRV
default features	Rust 1.51.0 (2021-03-25)
future	Rust 1.51.0 (2021-03-25)

It will keep a rolling MSRV policy of at least 6 months. If only the default features are enabled, MSRV will be updated conservatively. When using other features, like future, MSRV might be updated more frequently, up to the latest stable. In both cases, increasing MSRV is not considered a semver-breaking change.

Troubleshooting

Compile Errors on Some 32-bit Platforms

On some 32-bit target platforms including the followings, you may encounter compile errors:

• armv5te-unknown-linux-musleabi
• mips-unknown-linux-musl
• mipsel-unknown-linux-musl

error[E0432]: unresolved import `std::sync::atomic::AtomicU64`
  --> ... /moka-0.5.3/src/sync.rs:10:30
   |
10 |         atomic::{AtomicBool, AtomicU64, Ordering},
   |                              ^^^^^^^^^
   |                              |
   |                              no `AtomicU64` in `sync::atomic`

Such errors can occur because std::sync::atomic::AtomicU64 is not provided on these platforms but Moka uses it.

You can resolve the errors by disabling atomic64 feature, which is one of the default features of Moka. Edit your Cargo.toml to add default-features = false to the dependency declaration.

[dependencies]
moka = { version = "0.10", default-features = false }
# Or
moka = { version = "0.10", default-features = false, features = ["future"] }

This will make Moka to switch to a fall-back implementation, so it will compile.

Developing Moka

Running All Tests

To run all tests including future feature and doc tests on the README, use the following command:

$ RUSTFLAGS='--cfg skeptic --cfg trybuild' cargo test --all-features

Running All Tests without Default Features

$ RUSTFLAGS='--cfg skeptic --cfg trybuild' cargo test \
    --no-default-features --features future

Generating the Doc

$ cargo +nightly -Z unstable-options --config 'build.rustdocflags="--cfg docsrs"' \
    doc --no-deps --features 'future, dash'

Road Map

• async optimized caches. (v0.2.0)
• Size-aware eviction. (v0.7.0 via #24)
• API stabilization. (Smaller core cache API, shorter names for frequently used methods) (v0.8.0 via #105)
  • e.g.
  • get_or_insert_with(K, F) → get_with(K, F)
  • get_or_try_insert_with(K, F) → try_get_with(K, F)
  • blocking_insert(K, V) → blocking().insert(K, V)
  • time_to_live() → policy().time_to_live()
• Notifications on eviction. (v0.9.0 via #145)
• Cache statistics. (Hit rate, etc.)
• Upgrade TinyLFU to Window-TinyLFU. (details)
• The variable (per-entry) expiration, using a hierarchical timer wheel.

About the Name

Moka is named after the moka pot, a stove-top coffee maker that brews espresso-like coffee using boiling water pressurized by steam.

This name would imply the following facts and hopes:

• Moka is a part of the Java Caffeine cache family.
• It is written in Rust. (Many moka pots are made of aluminum alloy or stainless steel. We know they don't rust though)
• It should be fast. ("Espresso" in Italian means express)
• It should be easy to use, like a moka pot.

Credits

Caffeine

Moka's architecture is heavily inspired by the Caffeine library for Java. Thanks go to Ben Manes and all contributors of Caffeine.

cht

The source files of the concurrent hash table under moka::cht module were copied from the cht crate v0.4.1 and modified by us. We did so for better integration. cht v0.4.1 and earlier are licensed under the MIT license.

Thanks go to Gregory Meyer.

License

Moka is distributed under either of

• The MIT license
• The Apache License (Version 2.0)

at your option.

See LICENSE-MIT and LICENSE-APACHE for details.