Efficient way to ensure unique rows in SQLite3

Question

I am using SQLite3 in one of my projects and I need to ensure that the rows that are inserted into a table are unique with regard to a combination of some of their columns. In m

Answer 1

(I don't normally answer my own questions, but I would like to document a few ideas/partial solutions for this.)

The main problem with a composite primary key is the way the indexes are handled. Composite keys imply an index on the composite value, which in my case means indexing strings. While comparing string values isn't that slow, indexing a value with a length of, say, 500 bytes means that the B-tree nodes in the index can fit far fewer row/node pointers than a B-tree that indexes a 64-bit integer value. This means loading far more DB pages for each index search, as the height of the B-tree increases.

In order to deal with this issue I modified my code so that:

• It uses WAL mode. The performance increase was certainly worth such a small change, since I do not have any issues with the DB file not being self-contained.

• I used the MurmurHash3 hash function - after re-writing it in C and adapting it - to produce a single 32-bit hash value from the values of the fields that would form the key. I stored this hash in a new indexed column. Since this is an integer value, the index is quite fast. This is the only index for this table. Since there are going to be at most 10,000,000 rows in the table, hash collisions will not be a performance concern - although I cannot really consider the hash value to be UNIQUE, the index will only return a single row in the general case.

At this point there are two alternatives that I have coded and are currently undergoing testing:

• DELETE FROM Event WHERE Hash=? AND Fld0=? AND Fld2=? AND Fld3=?, followed by an INSERT.

• UPDATE Event SET Fld1=?,... WHERE Hash=? AND Fld0=? AND Fld2=? AND Fld3=?, followed by an INSERT if no rows where updated.

I expect the second alternative to be faster, but I will have to complete the testing first. In any case, it seems that with these changes the performance drop (as compared to the original index-less table) has been lessened to a factor of 5 or so, which is far more manageable.

EDIT:

At this point I have settled with using the second variation, which is indeed slightly faster. It seems, however, that any kind of index slows down SQLite3 dramatically as the indexed table gets larger. Increasing the DB page size to 8192 bytes seems to help somewhat, but not nearly as drastically as I would like.