mCaptcha-cache/README.md

mCaptcha Cache

Redis module that implements leaky bucket algorithm

Features

• Timers for individual count
• Clustering
• Persistence through RDB
• Persistence through AOF

Motivation

mCaptcha uses a leaky- bucket-enabled counter to keep track of traffic/challenge requests.

• At t=0(where t is time), if someone is visiting an mCaptcha-protected website, the counter for that website will be initialized and set to 1.

• It should also automatically decrement(by 1) after a certain period, say t=cooldown. We call this cool down period and is constant for a website.

• If at t=x(where x<cooldown), another user visits the same website, the counter becomes 2 and will auto decrement at t = cooldown + x for second user.

  Note that, for the decrement to work, we require two different timers that goes off at two different instants. The current(v0.1.3) of libmcaptcha implements this with internal data structures and timers --- something that can't be shared across several machines in a distributed setting.

  So we figured we'd use Redis to solve this problem and get synchronisation and persistence for free.

  This Redis module implements auto decrement on a special data type(which is also defined in this module).

How does it work?

If a timer is supposed to go off to decrement key myCounter at t=y(where y is an instant in future),

1. A hashmap called mcaptcha_cache:decrement:y(prefix might vary) is created with key-value pairs keyName: DecrementCount(myCounter: 1 in our case)

2. A timer will be created to go off at t=y

3. Any further decrement operations that are scheduled for t=y are registered with the same hashmap(mcaptcha_cache:decrement:y).

4. At t=y, a procedure will be executed to read all values of the hashmap(mcaptcha_cache:decrement:y) and performs all registered decrements. When its done, it cleans itself up.

This way, we are not spinning timers for every decrement operation but instead, one for every "pocket".

Gotchas:

This module creates and manages data of three types:

1. mcaptcha_cache:captcha:y where y(last character) is variable
2. mcaptcha_cache:pocket:x where x(last character) is variable
3. mcaptcha:timer:z where z(last character) is pocket name from step 2(See Hacks).

WARNING: Please don't modify these manually. If you do so, then Redis will panic

This module is capable of cleaning up after itself so manual clean up is unnecessary. If you have needs that are not met my this module and you which access/mutate data manually, please open an issue. I'd be happy to help.

Usage

There are two ways to run cache:

1. Using docker
2. On bare-metal

Docker

Use image from DockerHub:

$  docker run -p 6379:6379 mcaptcha/cache:0.1.1-alpha

or build from source:

Build

$ docker build -t mcaptcha/cache:latest .

Run

$  docker run -p 6379:6379 mcaptcha/cache:latest

Bare-metal

Build

Make sure you have Rust installed: https://www.rust-lang.org/tools/install

Then, build as usual:

cargo build --release

Run

redis-server --loadmodule ./target/release/libcache.so

Commands

Every counter has a name and a leak-rate in seconds.

Create/Increment counter

If counter exists, then count is incremented. Otherwise, it is created.

MCAPTCHA_CACHE.COUNT <counter-name> <leak-rate>

Benchmark

NOTE: These benchmarks are for reference only. Do not depend upon them too much. When in doubt, please craft and run benchmarks that are better suited to your workload.

To run benchmarks locally, launch Redis server with module loaded and:

$ ./scripts/bench.sh

• platform: Intel core i7-9750h

running set and get without pipelining
SET: 86095.57 requests per second, p50=0.311 msec
GET: 87519.70 requests per second, p50=0.311 msec

mCaptcha cache without piplining
MCAPTCHA_CACHE.COUNT mycounter 45: 85375.22 requests per second, p50=0.479 msec

running set and get with pipelining
SET: 822368.44 requests per second, p50=0.855 msec
GET: 900090.06 requests per second, p50=0.775 msec

mCaptcha cache with piplining
MCAPTCHA_CACHE.COUNT mycounter 45: 274876.31 requests per second, p50=2.767 msec

Hacks

I couldn't find any ways to persist timers to disk(RDB/AOF). So I'm using a dummy record(mcaptcha:timer:* see Gotchas) which will expire after an arbitrary time(see POCKET_EXPIRY_OFFSET in lib.rs). When that expiry occurs, I derive the key of the pocket from the values that are passed to expiration event handlers and perform clean up of both the pocket and counters registered with the pocket.

Ideally, I should be able to persist timers but I couldn't find ways to do that.