Bettering Concurrency in Redis Charge Limiting System

Background

Charge limiting is a way used to guard providers from overload. As well as, it may be used to forestall hunger of a multi-tenant useful resource by a number of very massive clients. At Rockset, we primarily use charge limiting to guard our:

1. metadata retailer from overload attributable to too many API requests.
2. log retailer from filling up as a consequence of mismatched enter and output charges
3. management airplane from too many state transitions.

We use Redisson RateLimiter which makes use of Redis underneath the hood to trace charge utilization. At a really primary degree, our utilization of the library seems like this (omitting particular enterprise logic for higher readability):

class RedisRateLimiter {
  non-public ultimate RRateLimiter rateLimitService = ...;

  public boolean isNotRateLimited(String key, int requestedTokens) {
      return rateLimitService.purchase(key, requestedTokens);
  }
}

Let’s not dive into the main points of RRateLimiter, however suffice it to say that this makes a community name to Redis. RedisRateLimiter.purchase will return true if requestedTokens wouldn’t exceed your charge restrict and false in any other case.

Downside

Lately, we noticed that as a consequence of many requests to Redis, the CPU on our Redis cluster was getting near 100%. The very first thing we tried was vertically scaling up our Redis occasion to purchase us time. Nonetheless, vertical scaling has its personal limits and each few weeks we’d find yourself with one other surge in Redis CPU.

We additionally observed that Redisson makes use of Lua scripting on the server facet and observed that lua compilation was taking on an honest chunk of CPU time. One other low hanging fruit we tried was configuring Redisson to cache lua compilation on the server facet, decreasing CPU time spent on this process. Since this was a easy config change, it didn’t require a code deploy and was straightforward to get out.

Aside from vertical scaling and enhancing configuration, we brainstormed a number of different approaches to the issue:

1. We may shard Redis over the speed restrict keys to unfold the load and horizontally scale.
2. We may queue charge restrict requests domestically and have a single thread that periodically (i.e. each 50ms) takes n gadgets off the queue and requests a bigger batch of tokens from Redis.
3. We may proactively reserve bigger batches of tokens and cache them domestically. When a request for tokens is available in, attempt coming back from the native cache. If that does not exist, go fetch a bigger batch. That is analogous to Malloc not making a sys name each time reminiscence is requested and as an alternative reserving bigger chunks that it manages.

Horizontally scaling Redis by sharding is a good long-term answer; it’s in all probability one thing we’re going to finish up doing sooner or later.

The issue with the second strategy is it raises a number of complexities: How often does the thread pull from the queue and ballot? Do you cap the scale of the queue and in that case, what occurs if the queue is full? How do you even set the cap on the queue? What if Redis has 50 tokens and we batch 10 requests every needing 10 tokens (asking Redis for a complete of 100 tokens)? Ideally 5 requests ought to succeed, however in actuality all 10 would fail. These issues are solvable, however would make the implementation fairly complicated. Thus, we ended up implementing the third answer.

As proven in direction of the tip of the submit, this implementation lowered Redis connections on charge restrict calls by 96%. The remainder of this submit will discover how we carried out the third strategy. It goes into a number of the pitfalls, complexities, and issues to contemplate when engaged on a batch-oriented answer resembling this one.

Implementation

Word that code offered on this weblog is in Java. Not all error dealing with is proven for simplicity. Additionally, I’ll reference a now() methodology which merely returns the unix timestamp in seconds from epoch.

Let’s begin easy:

class RedisRateLimiter {
  non-public ultimate RRateLimiter rateLimitService = ...;
  non-public ultimate lengthy batchSize = ...;
  non-public ultimate lengthy timeWindowSecs = ...;
  non-public lengthy reservedTokens = 0;
  non-public lengthy expirationTs = 0;

  public boolean isNotRateLimited(String key, int requestedTokens) {
    // On this case, we would as properly make a direct name to
    // simplify issues.
    if (requestedTokens > batchSize) {
      return rateLimitService.purchase(key, requestedTokens);
    }

    if (reservedTokens >= requestedTokens && expirationTs <= now()) {
      reservedTokens -= requestedTokens;
      return true;
    }

    if (rateLimitService.purchase(key, batchSize)) {
      reservedTokens = batchSize - requestedTokens;
      expirationTs = now() + timeWindowSecs;
      return true;
    }

    return false;
  }
}

This code seems high quality upon first look, however what occurs if a number of threads have to name isNotRateLimited on the identical time? The above code is actually not thread secure. I’ll go away as an train to the reader why making reservedTokens into an Atomic variable will not clear up the issue (though do tell us when you give you a intelligent lock-free answer). If Atomics will not work, we will attempt utilizing Locks as an alternative:

class RedisRateLimiter {
  non-public ultimate RRateLimiter rateLimitService = ...;
  non-public ultimate lengthy batchSize = ...;
  non-public ultimate lengthy timeWindowSecs = ...;
  non-public ultimate Lock lock = new ReentrantLock();
  non-public lengthy reservedTokens = 0;
  non-public lengthy expirationTs = 0;

  public boolean isNotRateLimited(String key, int requestedTokens) {
    // On this case, we would as properly make a direct name to
    // simplify issues.
    if (requestedTokens > batchSize) {
      return rateLimitService.purchase(key, requestedTokens);
    }

    lock.lock();
    attempt {
      if (reservedTokens >= requestedTokens && expirationTs <= now()) {
        reservedTokens -= requestedTokens;
        return true;
      } else if (expirationTs <= now()) {
        // Dissipate remaining tokens
        requestedTokens -= reservedTokens;
        reservedTokens = 0;
      }
    } lastly {
      // Straightforward to miss; do not lock throughout the community request.
      lock.unlock();
    }

    if (rateLimitService.purchase(key, batchSize)) {
      lock.lock();
      reservedTokens = (batchSize - requestedTokens);
      expirationTs = now() + timeWindowSecs;
      lock.unlock();
      return true;
    }

    return false;
  }
}

Whereas at first look this seems appropriate, there’s one refined drawback with it. What occurs if a number of threads see there aren’t sufficient reservedTokens? To illustrate reservedTokens is 0, our batchSize is 100, and 5 threads request 20 tokens every concurrently.

All 5 threads will see that there aren’t sufficient reserved tokens and every will fetch 100 tokens. Now, this machine is left with 450 reservedTokens and 5x too many requests to the exterior retailer. Can we do higher? All we actually want is for one thread to go and fetch a batch after which the opposite 4 threads can simply make the most of that batch. 1 community name, and fewer wasted tokens.

With some booleans and situation variables, we will fairly simply obtain this. In the event you’re unfamiliar with how situation variables work, try the java docs; most languages can have some form of situation variable implementation as properly. This is the code:

class RedisRateLimiter {
  non-public ultimate RRateLimiter rateLimitService = ...;
  non-public ultimate lengthy batchSize = ...;
  non-public ultimate lengthy timeWindowSecs = ...;
  non-public ultimate Lock lock = new ReentrantLock();
  non-public ultimate Situation fetchCondition = lock.newCondition();
  non-public boolean fetchInProgress = false;
  non-public lengthy reservedTokens = 0;
  non-public lengthy expirationTs = 0;

  public boolean isNotRateLimited(String key, int requestedTokens) {
    // On this case, we would as properly make a direct name to
    // simplify issues.
    if (requestedTokens > batchSize) {
      return rateLimitService.purchase(key, requestedTokens);
    }

    boolean doFetch = false;
    lock.lock();
    attempt {
      if (reservedTokens >= requestedTokens && expirationTs <= now()) {
        reservedTokens -= requestedTokens;
        return true;
      } else if (expirationTs <= now()) {
        requestedTokens -= reservedTokens;
        reservedTokens = 0;
      }

      if (fetchInProgress) {
        // Thread is already fetching; let's await it to complete.
        fetchCondition.await();
        if (reservedTokens >= requestedTokens) {
          reservedTokens -= requestedTokens;
          return true;
        }
        return false;
      } else {
        doFetch = true; // This thread ought to fetch the batch
        fetchInProgress = true; // Keep away from different threads from fetching.
      }
    } lastly {
      lock.unlock();
    }

    if (doFetch) {
      boolean acquired = rateLimitService.purchase(key, batchSize);
      lock.lock();
      if (acquired) {
        reservedTokens = (batchSize - requestedTokens);
        expirationTs = now() + timeWindowSecs;
      }
      fetchCondition.signalAll(); // Get up ready threads
      lock.unlock();
      return acquired;
    }

    return false;
  }
}

Now, we’ll solely ever have one thread at a time fetching a batch. Whereas the code is logically appropriate, we would find yourself charge limiting a thread too aggressively:

To illustrate our batch dimension is 100 and we’ve 5 threads requesting 25 tokens every concurrently. The primary thread (name it T1) will fetch the batch from the exterior service. The opposite 4 threads will wait on the situation variable. Nonetheless, the fifth thread can have waited for no purpose as a result of the primary 4 threads will burn up all of the tokens within the fetched batch. As a substitute, it may need been higher to both:

1. Instantly return false for the fifth thread (this may charge restrict too aggressively)
2. Or have the fifth thread make a direct name to the exterior service, not ready on the primary thread.

The second answer is carried out under:

class RedisRateLimiter {
  non-public ultimate RRateLimiter rateLimitService = ...;
  non-public ultimate lengthy batchSize = ...;
  non-public ultimate lengthy timeWindowSecs = ...;
  non-public ultimate Lock lock = new ReentrantLock();
  non-public ultimate Situation fetchCondition = lock.newCondition();
  non-public boolean fetchInProgress = false;
  non-public lengthy reservedTokens = 0;
  non-public lengthy expirationTs = 0;
  // Variety of tokens that ready threads will burn up.
  non-public lengthy unreservedFetchTokens = 0;
  // Utilized by ready threads to find out if the fetch they're
  // ready for succeeded or not.
  non-public boolean didFetchSucceed = false;

  public boolean isNotRateLimited(String key, int requestedTokens) {
    // On this case, we would as properly make a direct name to
    // simplify issues.
    if (requestedTokens > batchSize) {
      return rateLimitService.purchase(key, requestedTokens);
    }

    boolean doFetch = false;
    lock.lock();
    attempt {
      if (reservedTokens >= requestedTokens && expirationTimesatmp <= now()) {
        reservedTokens -= requestedTokens;
        return true;
      } else if (expirationTimestamp <= now()) {
        requestedTokens -= reservedTokens;
        reservedTokens = 0;        
      }

      if (fetchInProgress) {
        if (unreservedFetchTokens >= requestedTokens) {
          // Reserve your spot in line
          unreservedFetchTokens -= requestedTokens;
          fetchCondition.await();
          // If we get right here and the fetch succeeded, then we
          // are high quality.
          return didFetchSucceed;
        }
      } else {
        doFetch = true;
        fetchInProgress = true;
        unreservedFetchTokens = batch - requestedTokens;
      }
    } lastly {
      lock.unlock();
    }

    if (doFetch) {
      boolean acquired = rateLimitService.purchase(key, batchSize);
      lock.lock();
      didFetchSucceed = acquired;
      if (acquired) {
        reservedTokens = unreservedFetchTokens;
        expirationTs = now() + timeWindowSecs;
      }
      fetchCondition.signalAll(); // Get up ready threads
      lock.unlock();
      return acquired;
    }

    // If we get right here, it means there weren't sufficient
    // unreservedFetchTokens. Let's simply make our personal
    // name quite than ready in line.
    return rateLimitService.purchase(key, tokensRequested);
  }
}

Lastly, we have arrived at a suitable answer. In apply, the lock competition must be minimal as we’re solely setting a number of primitive values. However, as with something, it is best to benchmark this answer in your use case and see if it is sensible.

Setting the batch dimension

One remaining query is set batchSize. There’s a tradeoff right here: If batchSize is just too low, the variety of requests to Redis will strategy the variety of requests to isNotRateLimited. If batchSize is just too excessive, hosts will reserve too many tokens, ravenous out different hosts. One factor to contemplate is whether or not these hosts will be auto scaled. If that’s the case, as soon as numHosts * batchSize exceeds the speed restrict, different hosts will begin getting starved out even when the variety of requests is underneath the speed restrict.

To handle a few of this, it could be fascinating to discover utilizing a dynamically set batch dimension. If this machine used up the whole final batch, possibly it will possibly request 1.5x the batch subsequent time (with a cap after all). Alternatively, if batches are going to waste, maybe solely ask for half the batch subsequent time.

Outcomes

As an preliminary start line, we set the batchSize to be 1/1000 of the speed restrict for a given useful resource. For our workload, this resulted in ~4% of charge restrict requests going to Redis, an enormous enchancment. This may be seen within the chart under, the place the x-axis is time and the y-axis is p.c of requests hitting Redis:

Bettering our charge limiting at Rockset is an ongoing course of and this in all probability received’t be the final enchancment we have to make on this space. Keep tuned for extra. And when you’re serious about fixing these kind of issues, we’re hiring!

A fast apart

As an apart, the next code has a really refined concurrency bug. Can you see it?

class RedisRateLimiter {
  non-public ultimate RRateLimiter rateLimitService = ...;
  non-public ultimate lengthy batchSize = ...;
  non-public ultimate lengthy timeWindowSecs = ...;
  non-public ultimate Lock lock = new ReentrantLock();
  non-public ultimate Situation fetchCondition = lock.newCondition();
  non-public boolean fetchInProgress = false;
  non-public lengthy reservedTokens = 0;
  non-public lengthy expirationTs = 0;
  // Variety of tokens that ready threads will burn up.
  non-public lengthy unreservedFetchTokens = 0;

  public boolean isNotRateLimited(String key, int requestedTokens) {
    // On this case, we would as properly make a direct name to
    // simplify issues.
    if (requestedTokens > batchSize) {
      return rateLimitService.purchase(key, requestedTokens);
    }

    boolean doFetch = false;
    lock.lock();
    attempt {
      if (reservedTokens >= requestedTokens) {
        reservedTokens -= requestedTokens;
        return true;
      } else if (expirationTimestamp <= now()) {
        requestedTokens -= reservedTokens;
        reservedTokens = 0;        
      }

      if (fetchInProgress) {
        if (unreservedFetchTokens >= requestedTokens) {
          // Reserve your spot in line
          unreservedFetchTokens -= requestedTokens;
          fetchCondition.await();
          if (reservedTokens >= requestedTokens) {
            reservedTokens -= requestedTokens;
            return true;
          }
          return false;
        }
      } else {
        doFetch = true;
        fetchInProgress = true;
        unreservedFetchTokens = batch - requestedTokens;
      }
    } lastly {
      lock.unlock();
    }

    if (doFetch) {
      boolean acquired = rateLimitService.purchase(key, batchSize);
      lock.lock();
      if (acquired) {
        reservedTokens = (batchSize - requestedTokens);
        expirationTs = now() + timeWindowSecs;
      }
      fetchCondition.signalAll(); // Get up ready threads
      lock.unlock();
      return acquired;
    }

    // If we get right here, it means there weren't sufficient
    // unreservedFetchTokens. Let's simply make our personal
    // name quite than ready in line.
    return rateLimitService.purchase(key, tokensRequested);
  }
}

Trace: Even when rateLimitService.purchase all the time returned true, you may find yourself in conditions the place isNotRateLimited returns false.