Caching
The cache tier is a temporary data store layer, much faster than the database. The benefits of having a separate cache tier include better system performance, ability to reduce database workloads, and the ability to scale the cache tier independently
After receiving a request, a web server first checks if the cache has the available response. If it has, it sends data back to the client. If not, it queries the database, stores the response in cache, and sends it back to the client. This caching strategy is called a read-through cache. Other caching strategies are available depending on the data type, size, and access patterns. A previous study explains how different caching strategies work [6].
Interacting with cache servers is simple because most cache servers provide APIs for common programming languages. The following code snippet shows typical Memcached APIs:
SECONDS = 1
cache.set('myKey, 'hi there', 3600 * SECONDS)
cache.get('myKey')
Here are a few considerations for using a cache system:
Consider using cache when data is read frequently but modified infrequently. Since cached data is stored in volatile memory, a cache server is not ideal for persisting data. For instance, if a cache server restarts, all the data in memory is lost. Thus, important data should be saved in persistent data stores.
It is a good practice to implement an expiration policy. Once cached data is expired, it is removed from the cache. When there is no expiration policy, cached data will be stored in the memory permanently. It is advisable not to make the expiration date too short as this will cause the system to reload data from the database too frequently. Meanwhile, it is advisable not to make the expiration date too long as the data can become stale.
This involves keeping the data store and the cache in sync. Inconsistency can happen because data-modifying operations on the data store and cache are not in a single transaction. When scaling across multiple regions, maintaining consistency between the data store and cache is challenging. For further details, refer to the paper titled “Scaling Memcache at Facebook” published by Facebook [7].
A single cache server represents a potential single point of failure (SPOF), defined in Wikipedia as follows: “A single point of failure (SPOF) is a part of a system that, if it fails, will stop the entire system from working” [8]. As a result, multiple cache servers across different data centers are recommended to avoid SPOF. Another recommended approach is to overprovision the required memory by certain percentages. This provides a buffer as the memory usage increases.
Once the cache is full, any requests to add items to the cache might cause existing items to be removed. This is called cache eviction. Least-recently-used (LRU) is the most popular cache eviction policy. Other eviction policies, such as the Least Frequently Used (LFU) or First in First Out (FIFO), can be adopted to satisfy different use cases.
This is perhaps the most commonly used caching approach, at least in the projects that I worked on. The cache sits on the side and the application directly talks to both the cache and the database. There is no connection between the cache and the primary database. All operations to cache and the database are handled by the application.
- The application first checks the cache.
- If the data is found in cache, we’ve cache hit. The data is read and returned to the client.
- If the data is not found in cache, we’ve cache miss. The application has to do some extra work. It queries the database to read the data, returns it to the client and stores the data in cache so the subsequent reads for the same data results in a cache hit.
Cache-aside caches are usually general purpose and work best for read-heavy workloads. Memcached and Redis are widely used. Systems using cache-aside are resilient to cache failures. If the cache cluster goes down, the system can still operate by going directly to the database. (Although, it doesn’t help much if cache goes down during peak load. Response times can become terrible and in worst case, the database can stop working.)
Another benefit is that the data model in cache can be different than the data model in database. E.g. the response generated as a result of multiple queries can be stored against some request id.
When cache-aside is used, the most common write strategy is to write data to the database directly. When this happens, cache may become inconsistent with the database. To deal with this, developers generally use time to live (TTL) and continue serving stale data until TTL expires. If data freshness must be guaranteed, developers either invalidate the cache entry or use an appropriate write strategy, as we’ll explore later.
Read-through cache sits in-line with the database. When there is a cache miss, it loads missing data from database, populates the cache and returns it to the application.
Both cache-aside and read-through strategies load data lazily, that is, only when it is first read.
While read-through and cache-aside are very similar, there are at least two key differences:
- In cache-aside, the application is responsible for fetching data from the database and populating the cache. In read-through, this logic is usually supported by the library or stand-alone cache provider.
- Unlike cache-aside, the data model in read-through cache cannot be different than that of the database.
Read-through caches work best for read-heavy workloads when the same data is requested many times. For example, a news story. The disadvantage is that when the data is requested the first time, it always results in cache miss and incurs the extra penalty of loading data to the cache. Developers deal with this by ‘warming’ or ‘pre-heating’ the cache by issuing queries manually. Just like cache-aside, it is also possible for data to become inconsistent between cache and the database, and solution lies in the write strategy, as we’ll see next.
- The application writes the data directly to the cache.
- The cache updates the data in the main database. When the write is complete, both the cache and the database have the same value and the cache always remains consistent.
On its own, write-through caches don’t seem to do much, in fact, they introduce extra write latency because data is written to the cache first and then to the main database (two write operations.) But when paired with read-through caches, we get all the benefits of read-through and we also get data consistency guarantee, freeing us from using cache invalidation (assuming ALL writes to the database go through the cache.)
DynamoDB Accelerator (DAX) is a good example of read-through / write-through cache. It sits inline with DynamoDB and your application. Reads and writes to DynamoDB can be done through DAX. (Side note: If you are planning to use DAX, please make sure you familiarize yourself with its data consistency model and how it interplays with DynamoDB.)
Here, data is written directly to the database and only the data that is read makes it way into the cache.
Write-around can be combine with read-through and provides good performance in situations where data is written once and read less frequently or never. For example, real-time logs or chatroom messages. Likewise, this pattern can be combined with cache-aside as well.
Here, the application writes data to the cache which stores the data and acknowledges to the application immediately. Then later, the cache writes the data back to the database.
This is very similar to to Write-Through but there’s one crucial difference: In Write-Through, the data written to the cache is synchronously updated in the main database. In Write-Back, the data written to the cache is asynchronously updated in the main database. From the application perspective, writes to Write-Back caches are faster because only the cache needed to be updated before returning a response.
Write back caches improve the write performance and are good for write-heavy workloads. When combined with read-through, it works good for mixed workloads, where the most recently updated and accessed data is always available in cache.
It’s resilient to database failures and can tolerate some database downtime. If batching or coalescing is supported, it can reduce overall writes to the database, which decreases the load and reduces costs, if the database provider charges by number of requests e.g. DynamoDB. Keep in mind that DAX is write-through so you won’t see any reductions in costs if your application is write heavy. (When I first heard of DAX, this was my first question - DynamoDB can be very expensive, but damn you Amazon.)
Some developers use Redis for both cache-aside and write-back to better absorb spikes during peak load. The main disadvantage is that if there’s a cache failure, the data may be permanently lost.
Most relational databases storage engines (i.e. InnoDB) have write-back cache enabled by default in their internals. Queries are first written to memory and eventually flushed to the disk.