• 9 July 2012

Redis Internals

Everything in Redis is ultimately represented as a string. Even collections like lists, sets, sorted sets, and maps are composed of strings. Redis defines a special structure, which it calls simple dynamic string or SDS. This structure consists of three parts, namely:

  • buff — A character array that stores the string
  • len — A long type that stores the length of the buff array
  • free — Number of additional bytes available for use

Although you may think of storing len separately as an overhead, because it can be easily calculated based on the buff array, it allows for string length lookup in fixed time.

Redis keeps its data set in the primary memory, persisting it to disk as required. Unlike MongoDB, it does not use memory-mapped files for that purpose. Instead, Redis implements its own virtual memory subsystem. When a value is swapped to disk, a pointer to that disk page is stored with the key. Read more about the virtual memory technical specification at http://code.google.com/p/redis/wiki/VirtualMemorySpecification.

In addition to the virtual memory manager, Redis also includes an event library that helps coordinate the non-blocking socket operations.


  • Redis doesn’t rely on operating system swapping because:
    Redis objects don’t map one-to-one with swap pages. Swap pages are 4,096 bytes long and Redis objects could span more than one page. Similarly, more than one Redis object could be in a single swap page. Therefore, even when a small percentage of the Redis objects are accessed, it’s possible a large number of swap pages are touched. Operating system swapping keeps track of swap page access. Therefore, even if a byte in a swap page is accessed it is left out by the swapping system.
  • Unlike MongoDB, Redis data format when in RAM and in disk are not similar. Data on disk is compressed way more as compared to its RAM counterpart. Therefore, using custom swapping involves less disk I/O.

Salvatore Sanfillipo, the creator of Redis, talks about the Redis virtual memory system in his blog post titled, “Redis Virtual Memory: the story and the code,” at http://antirez.com/post/redis-virtual-memory-story.html.

  • 6 July 2012

NoSql Notes

In NoSQL stores, the create and read operations are more important than the update and delete operations, so much so that sometimes those are the only operations.