query_cache

#Caching

##Overview The plugin query cache maintains a bidirectional GRPC stream connection to the cache server (hosted by the plugin manager process)

Cache data is streamed row by row, in both directions.

###Cache Set

• call CacheCommand_SET_RESULT_START command
• for each row:
  • stream to FDW
  • stream to cache with CacheCommand_SET_RESULT_ITERATE command
• mark completion with CacheCommand_SET_RESULT_END
• fetch index bucket for this table (if it exists): CacheCommand_GET_INDEX
• update index bucket and write back CacheCommand_SET_INDEX

###Cache Get

• fetch index bucket for this table (if it exists): CacheCommand_GET_INDEX
• if this is a cache hit (i.e. an index item satisfying columns and quals exists), call CacheCommand_GET_RESULT
• cache will stream results to us a row at a time - for each row
  • stream to fdw

##Concurrency The plugin may make more than one request to the cache at a time (for example, if a query consists of multiple scans with the same connetion, i.e. multiple concurrent plugin Execution calls).

Therefore results from the cache may be interleaved, e.g. the results from 2 different cache gets may be streaming at the same time

To handle this, each cache command is made using a CallId (unique time-based identifier). When making a cache command, the query cache registers a callback which is stored in a map keyed by call id.

When the cache stream receiver processes a response from the cache, it uses the response callId to find the appropriate callback function and invokes it.

##TTL

The TTL is controlled by the client, as multiple clients might be using the same connections and each may have a different TTL.

This is achieved as follows. The cache saves data with a default TTL of 5 minutes (this is the TTL passed to the underlying cache). When fetching data, the age of the retrieved data is determined using the insertion time. If the age of the data is older than the client TTL, it is considered a cache miss

NOTE: If the client TTL is greater than the default TTL of the cache, the default TTL is increased as appropriate to ensure the data remains in the cache at lkeast as long as is needed for the client TTL