mysql innodb buffer vs memory table performance

Question

Is a memory/heap engine table the same performance wise to a mostly innodb table database with big buffer pool? I usually have 2 tables - 1 innodb with varchars and several rows and a memory table compact size (5 rows, mostly just PK and indexed ints for heavy reads..I recently learned about innodb buffer so is my table clone system overkill and useless or still faster then innodb?

Answer 1

In memory tables must be more performant, at least in theory: in InnoDB, even with a large buffer pool, you're going to have block-based structure in the cache, so some blocks will only be partially full, and that's an overhead. Another reason is that in-memory tables don't have row versions or row locks, so, again, this is going to use less memory. But befare: in-memory tables still don't have row-level locking, so if you run large updates, you may actually find that using InnoDB is more scalable.

So, to sum up: MEMORY table - potentially less memory to store the same amount of data, InnoDB - potentially more scalable.

Everything needs to be measured for your particular case of course.

Perhaps if you need to store data in memory anyway, choose an in-memory database? (shameless plug).

Answer 2

Reads from the InnoDB buffer pool will be sensibly as fast as with Memory tables.

In some cases, Memory tables could even out-perform buffered InnoDB tables, the former also supports Hash indexes whereas the latter only supports B-Tree indexes. Depending on the profile of your queries, you might get faster reads with Hash tables.

Besides, buffered InnoDB tables could be flushed out of the buffer if some query require this memory space, or if the data is seldom used. By explicitely copying your data to a Memory table, you have the guarantee that your data will always be in memory.

I should also mention that regardless of the size of the buffer pool, updates to an InnoDB table will need to be flushed to disk at some stage. But I understand this does not apply in your use case.

Now this is theory. Only if this data is to be read very, very frequently should you bother with these considerations.