How do I link records to a large table efficiently using python Dedupe?

Question

I'm trying to use the Dedupe package to merge a small messy data to a canonical table. Since the canonical table is very large (122 million rows), I can't load it all into memory.

The current approach that I'm using based off this takes an entire day on test data: a 300k row table of messy data stored in a dict, and a 600k row table of canonical data stored in mysql. If I do it all in memory (read the canonical table in as a dict) it only takes half an hour.

Is there a way to make this more efficient?

blocked_pairs = block_data(messy_data, canonical_db_cursor, gazetteer)
clustered_dupes = gazetteer.matchBlocks(blocked_pairs, 0)

def block_data(messy_data, c, gazetteer):

    block_groups = itertools.groupby(gazetteer.blocker(messy_data.viewitems()),
                                     lambda x: x[1])
    for (record_id, block_keys) in block_groups:

        a = [(record_id, messy_data[record_id], set())]

        c.execute("""SELECT *
                    FROM canonical_table
                    WHERE record_id IN
                        (SELECT DISTINCT record_id
                         FROM blocking_map
                         WHERE block_key IN %s)""", 
                  (tuple(block_key for block_key, _ in block_keys),))

        b = [(row[self.key], row, set()) for row in c]

        if b:
            yield (a, b)

Reuter

Answer 1

Sped it up dramatically by splitting up the query into two queries. I'm using mysql and all the columns used in the example are indexed...

def block_data(messy_data, c, gazetteer):

    block_groups = itertools.groupby(gazetteer.blocker(messy_data.viewitems()),
                                 lambda x: x[1])
    for (record_id, block_keys) in block_groups:

        a = [(record_id, messy_data[record_id], set())]

        c.execute("""SELECT DISTINCT record_id
                     FROM blocking_map
                     WHERE block_key IN %s""", 
                  (tuple(block_key for block_key, _ in block_keys),))

        values = tuple(row['record_id'] for row in c)

        if values:

            c.execute("""SELECT *
                         FROM canonical_table
                         WHERE record_id IN %s""",
                      (values,))

            b = [(row['record_id'], row, set())
                 for row in c]

            if b:
                yield (a, b)

Reuter

Answer 2

It would probably be faster yet if you expressed the query as a JOIN:

SELECT canonical_table.*
FROM canonical_table
JOIN blocking_map 
ON (canonical_table.record_id = blocking_map.record_id)
WHERE blocking_map IN %s

Your solution is basically doing a join in Python so probably the database would do it better. The "IN" syntax in your original attempt is rarely optimized as well as a correct join.

Humo