EmergencyEMERGENCY? Get 24/7 Help Now!

Understanding GCache and Record-Set Cache in Percona XtraDB Cluster

 | April 11, 2016 |  Posted In: MySQL, Percona XtraDB Cluster

PREVIOUS POST
NEXT POST

In this blog, we will examine the differences between GCache and Record-Set Cache in Percona XtraDB Cluster.

In Percona XtraDB Cluster (PXC), there is the concept of GCache and Record-Set cache (which can also be called transaction write-set cache). The use of these two caches is often confusing if you are running long transactions, as both of them result in the creation of disk-level files. Let’s understand what their main differences are.

Record-Set Cache
  • When you run a long-running transaction on any particular node, it will try to append a key for each row that it tries to modify (the key is a unique identifier for the row {db,table,pk.columns}). This information is cached in out-write-set, which is then sent to the group for certification. To start with, keys are cached in HeapStore (which has page-size=65K and total-size=4MB). If the transaction data-size outgrows this limit, then the storage is switched from Heap to Page (which has a page-size=64MB and total-limit=free-space-on-disk). All these limits are non-configurable, but having a memory-page size greater than 4MB per transaction can cause things to stall due to memory pressure, so this limit is reasonable. (This is another limitation to address when Galera supports large transaction.)
  • The same long-running transaction will also generate binlog data that also appends to out-write-set on commit using the same technique explained above (HeapStore->FileStore). This data could be significant as it is a binlog image of rows inserted/updated/deleted by the transaction. wsrep_max_ws_size controls the size of this part of the write set. (The threshold doesn’t consider size allocated for caching-keys (above) and the header).
  • If FileStore is used, it creates a file on the disk (with names like xxxx_keys and xxxx_data) to store the cache data. These files are kept until a transaction is committed, so the lifetime of the transaction is linked.
  • When the node is done with the transaction and is about to commit, it will generate the final-write-set using the two files (if the data size grew enough to use FileStore) plus  HEADER, and will publish it for certification to cluster.
  • The native node executing the transaction will also act as subscription node, and will receive its own write-set through the cluster publish mechanism. This time, the native node will try to cache write-set into its GCache. How much data GCache retains is controlled by the GCache configuration.

GCache:

  • GCache holds the write-set published on the cluster for replication.
  • The lifetime of write-set in GCache is not transaction linked.
  • When a JOINER node needs an IST, it will be serviced through this GCache (if possible).
  • GCache will also create the files to disk. (You can read more about it here.)

Interestingly, at any given point in time, the native node has two copies of the write-set: one in GCache and other in Record-Set Cache.

For example:

I tried inserting/updating 2M rows in a table with a schema like:

and it created write-set key/data files in the background that looked like this

PREVIOUS POST
NEXT POST
Krunal Bauskar

Krunal is PXC lead at Percona. He is responsible for day-day PXC development, what goes into PXC, bug fixes, releases, etc.. Before joining Percona he use to work as part of InnoDB team at MySQL/Oracle. He authored most of the temporary table revamp work, undo log truncate, atomic truncate and lot of other features. In past he was associated with Yahoo! Labs researching on bigdata problems and database startup which is now part of Teradata. His interest mainly includes data-management at any scale and has been practicing it for more than decade now.

One Comment

  • There is small typo
    “HeapStore (which has page-size=65K and total-size=4MB)”
    should be
    “HeapStore (which has page-size=*64K* and total-size=4MB)”

Leave a Reply