My SQL query can be very simply written as:
result = session.query(Table).filter(Table.my_key._in(key_set))
The my_key integer column is indexed (primary key), but key_set could be very large indeed, with tens of millions of values.
What's the recommended SQLAlchemy pattern for filtering by such huge sets?
Is there something built-in that's more efficient than a pedestrian:
result = [session.query(Table).get(key) for key in key_set]
In such an extreme case you're better off thinking what is the recommended SQL solution first, and then implementing that in SQLAlchemy – even using raw SQL, if need be. One such solution is to create a temporary table for key_set data and to populate it.
In order to test something like your setup, I created the following model
class Table(Base):
__tablename__ = 'mytable'
my_key = Column(Integer, primary_key=True)
and populated it with 20,000,000 rows:
In [1]: engine.execute("""
...: insert into mytable
...: select generate_series(1, 20000001)
...: """)
I also created some helpers for testing different combinations of temporary tables, populating, and queries. Note that the queries use the Core table, in order to bypass the ORM and its machinery – the contribution to timings would be constant anyway:
# testdb is just your usual SQLAlchemy imports, and some
# preconfigured engine options.
from testdb import *
from sqlalchemy.ext.compiler import compiles
from sqlalchemy.sql.expression import Executable, ClauseElement
from io import StringIO
from itertools import product
class Table(Base):
__tablename__ = "mytable"
my_key = Column(Integer, primary_key=True)
def with_session(f):
def wrapper(*a, **kw):
session = Session(bind=engine)
try:
return f(session, *a, **kw)
finally:
session.close()
return wrapper
def all(_, query):
return query.all()
def explain(analyze=False):
def cont(session, query):
results = session.execute(Explain(query.statement, analyze))
return [l for l, in results]
return cont
class Explain(Executable, ClauseElement):
def __init__(self, stmt, analyze=False):
self.stmt = stmt
self.analyze = analyze
@compiles(Explain)
def visit_explain(element, compiler, **kw):
stmt = "EXPLAIN "
if element.analyze:
stmt += "ANALYZE "
stmt += compiler.process(element.stmt, **kw)
return stmt
def create_tmp_tbl_w_insert(session, key_set, unique=False):
session.execute("CREATE TEMPORARY TABLE x (k INTEGER NOT NULL)")
x = table("x", column("k"))
session.execute(x.insert().values([(k,) for k in key_set]))
if unique:
session.execute("CREATE UNIQUE INDEX ON x (k)")
session.execute("ANALYZE x")
return x
def create_tmp_tbl_w_copy(session, key_set, unique=False):
session.execute("CREATE TEMPORARY TABLE x (k INTEGER NOT NULL)")
# This assumes that the string representation of the Python values
# is a valid representation for Postgresql as well. If this is not
# the case, `cur.mogrify()` should be used.
file = StringIO("".join([f"{k}\n" for k in key_set]))
# HACK ALERT, get the DB-API connection object
with session.connection().connection.connection.cursor() as cur:
cur.copy_from(file, "x")
if unique:
session.execute("CREATE UNIQUE INDEX ON x (k)")
session.execute("ANALYZE x")
return table("x", column("k"))
tmp_tbl_factories = {
"insert": create_tmp_tbl_w_insert,
"insert (uniq)": lambda session, key_set: create_tmp_tbl_w_insert(session, key_set, unique=True),
"copy": create_tmp_tbl_w_copy,
"copy (uniq)": lambda session, key_set: create_tmp_tbl_w_copy(session, key_set, unique=True),
}
query_factories = {
"in": lambda session, _, x: session.query(Table.__table__).
filter(Table.my_key.in_(x.select().as_scalar())),
"exists": lambda session, _, x: session.query(Table.__table__).
filter(exists().where(x.c.k == Table.my_key)),
"join": lambda session, _, x: session.query(Table.__table__).
join(x, x.c.k == Table.my_key)
}
tests = {
"test in": (
lambda _s, _ks: None,
lambda session, key_set, _: session.query(Table.__table__).
filter(Table.my_key.in_(key_set))
),
"test in expanding": (
lambda _s, _kw: None,
lambda session, key_set, _: session.query(Table.__table__).
filter(Table.my_key.in_(bindparam('key_set', key_set, expanding=True)))
),
**{
f"test {ql} w/ {tl}": (tf, qf)
for (tl, tf), (ql, qf)
in product(tmp_tbl_factories.items(), query_factories.items())
}
}
@with_session
def run_test(session, key_set, tmp_tbl_factory, query_factory, *, cont=all):
x = tmp_tbl_factory(session, key_set)
return cont(session, query_factory(session, key_set, x))
For small key sets the simple IN query you have is about as fast as the others, but using a key_set of 100,000 the more involved solutions start winning:
In [10]: for test, steps in tests.items():
...: print(f"{test:<28}", end=" ")
...: %timeit -r2 -n2 run_test(range(100000), *steps)
...:
test in 2.21 s ± 7.31 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test in expanding 630 ms ± 929 µs per loop (mean ± std. dev. of 2 runs, 2 loops each)
test in w/ insert 1.83 s ± 3.73 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test exists w/ insert 1.83 s ± 3.99 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test join w/ insert 1.86 s ± 3.76 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test in w/ insert (uniq) 1.87 s ± 6.67 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test exists w/ insert (uniq) 1.84 s ± 125 µs per loop (mean ± std. dev. of 2 runs, 2 loops each)
test join w/ insert (uniq) 1.85 s ± 2.8 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test in w/ copy 246 ms ± 1.18 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test exists w/ copy 243 ms ± 2.31 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test join w/ copy 258 ms ± 3.05 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test in w/ copy (uniq) 261 ms ± 1.39 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test exists w/ copy (uniq) 267 ms ± 8.24 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
test join w/ copy (uniq) 264 ms ± 1.16 ms per loop (mean ± std. dev. of 2 runs, 2 loops each)
Raising the key_set to 1,000,000:
In [11]: for test, steps in tests.items():
...: print(f"{test:<28}", end=" ")
...: %timeit -r2 -n1 run_test(range(1000000), *steps)
...:
test in 23.8 s ± 158 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test in expanding 6.96 s ± 3.02 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test in w/ insert 19.6 s ± 79.3 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test exists w/ insert 20.1 s ± 114 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test join w/ insert 19.5 s ± 7.93 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test in w/ insert (uniq) 19.5 s ± 45.4 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test exists w/ insert (uniq) 19.6 s ± 73.6 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test join w/ insert (uniq) 20 s ± 57.5 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test in w/ copy 2.53 s ± 49.9 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test exists w/ copy 2.56 s ± 1.96 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test join w/ copy 2.61 s ± 26.8 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test in w/ copy (uniq) 2.63 s ± 3.79 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
test exists w/ copy (uniq) 2.61 s ± 916 µs per loop (mean ± std. dev. of 2 runs, 1 loop each)
test join w/ copy (uniq) 2.6 s ± 5.31 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)
Key set of 10,000,000, COPY solutions only, since the others ate all my RAM and were going through swap before killed, hinting that they'd never finish on this machine:
In [12]: for test, steps in tests.items():
...: if "copy" in test:
...: print(f"{test:<28}", end=" ")
...: %timeit -r1 -n1 run_test(range(10000000), *steps)
...:
test in w/ copy 28.9 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
test exists w/ copy 29.3 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
test join w/ copy 29.7 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
test in w/ copy (uniq) 28.3 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
test exists w/ copy (uniq) 27.5 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
test join w/ copy (uniq) 28.4 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)
So, for small key sets (~100,000 or less) it doesn't matter that much what you use, though using expanding bindparam is a clear winner in time compared to ease of use, but for much larger sets you might want to consider using a temporary table and COPY.
It is notable that for large sets the query plans are identical, if using the unique index:
In [13]: print(*run_test(range(10000000),
...: tmp_tbl_factories["copy (uniq)"],
...: query_factories["in"],
...: cont=explain()), sep="\n")
Merge Join (cost=45.44..760102.11 rows=9999977 width=4)
Merge Cond: (mytable.my_key = x.k)
-> Index Only Scan using mytable_pkey on mytable (cost=0.44..607856.88 rows=20000096 width=4)
-> Index Only Scan using x_k_idx on x (cost=0.43..303939.09 rows=9999977 width=4)
In [14]: print(*run_test(range(10000000),
...: tmp_tbl_factories["copy (uniq)"],
...: query_factories["exists"],
...: cont=explain()), sep="\n")
Merge Join (cost=44.29..760123.36 rows=9999977 width=4)
Merge Cond: (mytable.my_key = x.k)
-> Index Only Scan using mytable_pkey on mytable (cost=0.44..607856.88 rows=20000096 width=4)
-> Index Only Scan using x_k_idx on x (cost=0.43..303939.09 rows=9999977 width=4)
In [15]: print(*run_test(range(10000000),
...: tmp_tbl_factories["copy (uniq)"],
...: query_factories["join"],
...: cont=explain()), sep="\n")
Merge Join (cost=39.06..760113.29 rows=9999977 width=4)
Merge Cond: (mytable.my_key = x.k)
-> Index Only Scan using mytable_pkey on mytable (cost=0.44..607856.88 rows=20000096 width=4)
-> Index Only Scan using x_k_idx on x (cost=0.43..303939.09 rows=9999977 width=4)
Since the test tables are sort of artificial, it is able to use index only scans.
Finally, here are the timings for the "pedestrian" method, for a rough comparison:
In [3]: for ksl in [100000, 1000000]:
...: %time [session.query(Table).get(k) for k in range(ksl)]
...: session.rollback()
...:
CPU times: user 1min, sys: 1.76 s, total: 1min 1s
Wall time: 1min 13s
CPU times: user 9min 48s, sys: 17.3 s, total: 10min 5s
Wall time: 12min 1s
The problem is that using Query.get() necessarily includes the ORM, while the original comparisons did not. Still, it should be somewhat obvious that the separate roundtrips to the database cost dearly, even when using a local database.