Repository navigation
DuckDB plans views that share reads in time exponential in their depth — draft for upstream #292
Description
🌀 Claude: DuckDB's planning time grows exponentially when views read the previous view several times, although the plan it finds stays small. The GraphCast demo (#270) writes each processor round as a view of the round before. On it, DuckDB plans one round in 45 s, and doesn't finish planning two within 10 minutes. This is a draft for an upstream DuckDB issue.
Reproduction
Each view reads the one before three times: as each edge's sender, as its receiver, and as a residual. This is one round of message passing in a graph neural network, with a layer norm's windows. The script needs only duckdb, and runs in seconds up to round 5:
import sys
import time
import duckdb
nodes, edges, f = 100, 400, 4
con = duckdb.connect()
con.execute(f"CREATE TABLE node AS SELECT i // {f} AS id, i % {f} AS f, random() AS val FROM range({nodes * f}) t(i)")
con.execute(f"CREATE TABLE edge AS SELECT i AS edge, (i * 7) % {nodes} AS sender, (i * 13) % {nodes} AS receiver FROM range({edges}) t(i)")
con.execute(f"CREATE TABLE w AS SELECT i // {2*f} AS f, i % {2*f} AS g, random() AS w FROM range({2 * f * 2 * f}) t(i)")
con.execute("CREATE VIEW mesh_0 AS SELECT * FROM node")
print("duckdb", duckdb.__version__)
for r in range(int(sys.argv[1])):
con.execute(f"""CREATE VIEW mesh_{r + 1} AS
WITH edge_in AS (
SELECT e.edge AS id, s.f, s.val FROM edge e JOIN mesh_{r} s ON e.sender = s.id
UNION ALL
SELECT e.edge AS id, x.f + {f} AS f, x.val FROM edge e JOIN mesh_{r} x ON e.receiver = x.id),
message AS (
SELECT i.id, w.g AS f, SUM(i.val * w.w) AS val
FROM edge_in i JOIN w ON i.f = w.f GROUP BY i.id, w.g),
normed AS (
SELECT id, f, (val - AVG(val) OVER (PARTITION BY id))
/ sqrt(VAR_POP(val) OVER (PARTITION BY id) + 1e-5) AS val
FROM message),
summed AS (
SELECT e.receiver AS id, m.f, SUM(m.val) AS val
FROM normed m JOIN edge e ON m.id = e.edge WHERE m.f < {f} GROUP BY e.receiver, m.f)
SELECT x.id, x.f, x.val + coalesce(s.val, 0) AS val
FROM mesh_{r} x LEFT JOIN summed s ON x.id = s.id AND x.f = s.f""")
start = time.perf_counter()
plan = con.sql(f"EXPLAIN SELECT * FROM mesh_{r + 1}").fetchall()[0][1]
print(f"mesh_{r + 1}: planned in {time.perf_counter() - start:6.2f} s, {plan.count('┌')} boxes", flush=True)python view_plan_repro.py 6, with DuckDB 1.5.6 on an Intel i7-8700:
| View | Planning (EXPLAIN) |
Boxes in the plan |
|---|---|---|
mesh_1 |
0.00 s | 28 |
mesh_2 |
0.01 s | 57 |
mesh_3 |
0.05 s | 79 |
mesh_4 |
0.45 s | 81 |
mesh_5 |
6.71 s | 83 |
mesh_6 |
111 s | 85 |
Planning time grows about 15× per view, while the plan grows by two boxes. So DuckDB finds that each view's reads are shared, and computes them once, but only after doing work exponential in the depth. It looks as if the views are expanded inline before the shared reads are found.
In the GraphCast demo
The same shape, with GraphCast_small's real tables: 308 operators for the encoder (mesh_0), planned in 0.21 s; 504 for one processor round (mesh_1), planned in 45 s; mesh_2 didn't finish planning in 10 minutes. A forecast has 16 rounds. DataFusion doesn't plan it either, for a different reason: it inlines every read, so its plan grows 4.7× per round (#267).
For upstream
- Search DuckDB's issues for planning time with nested views or CTEs, before filing.
- If there's no existing issue, file this script and its table.
- The demo keeps its SQL as written; it will run on DuckDB once planning is linear in depth.
Activity