Joining with a dj.U restriction returns duplicate rows and inflates aggr counts: the DISTINCT is dropped

Đang mở
#1,660 0 bình luận 0 reaction 0 người được giao Xem trên GitHub

Maintainer thường phản hồi trong vòng 6 ngày

@dimitri-yatsenko đang làm issue này rồi.

Từ ngày 7/11/2026.

  • #1662 của @dimitri-yatsenko — đang mở

Đánh giá

Độ khó
4/5
Thời gian dự kiến
3-5 ngày
Mức phù hợp với người mới
65/100
Loại issue
Lỗi
Độ rõ ràng
Đặc tả rõ ràng
Mức độ hoạt động
Sôi nổi
Công nghệ
mysql, postgresql, python
Lĩnh vực
backend, databases

Hướng nghiên cứu

Start with src/datajoint/expression.py, reading QueryExpression.join and dj.U.aggr, where the issue identifies _top handling as a nearby precedent for wrapping operands. Check the existing tests for dj.U, joins, aggr, and grouping, then reproduce the listed cases on MySQL or PostgreSQL. Done means distinct results and aggregate counts match the expected values for both operand positions and dj.U.aggr, with existing tests still passing.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.

Mô tả

bug

What happens

dj.U("parent_id") & Child lists each parent_id once: its SQL is SELECT DISTINCT parent_id FROM child. When this expression is used inside another query (a join on either side, extend, aggr, or dj.U(...).aggr), the DISTINCT is dropped. The operand is joined as the plain child table, so it contributes one row per child instead of one row per parent. Joins return repeated rows and aggr counts are too high. Nothing warns.

len() of the same join is correct, because it counts distinct primary-key values. So len(q) and len(q.to_dicts()) disagree. The preview shows four rows above (Total: 2).

Restricting by the expression works: Parent & (dj.U("parent_id") & Child) and Parent - (dj.U("parent_id") & Child) return the right parents.

Reproduction

On master (3c912c46), the results are the same on MySQL 8.0 and PostgreSQL 15:

import datajoint as dj

schema = dj.Schema("u_distinct_repro")

@schema
class Parent(dj.Manual):
    definition = """
    parent_id : int32
    ---
    label : varchar(8)
    """

@schema
class Child(dj.Manual):
    definition = """
    -> Parent
    child_id : int32
    ---
    color : varchar(8)
    """

Parent.insert([{"parent_id": 1, "label": "a"}, {"parent_id": 2, "label": "b"}, {"parent_id": 3, "label": "a"}])
Child.insert([
    {"parent_id": 1, "child_id": 1, "color": "red"},
    {"parent_id": 1, "child_id": 2, "color": "blue"},
    {"parent_id": 2, "child_id": 1, "color": "blue"},
    {"parent_id": 2, "child_id": 2, "color": "blue"},
])

has_children = dj.U("parent_id") & Child         # alone: parents 1, 2
colors = dj.U("parent_id", "color") & Child      # alone: (1, blue), (1, red), (2, blue)
Query Expected Actual
Parent * has_children (1, a), (2, b) (1, a), (1, a), (2, b), (2, b)
has_children * Parent (1, a), (2, b) (1, a), (1, a), (2, b), (2, b)
(has_children & "parent_id < 3") * Parent (1, a), (2, b) (1, a), (1, a), (2, b), (2, b)
Parent * colors (1, blue), (1, red), (2, blue) (2, blue) appears twice
Parent.extend(has_children) parents 1, 2, 3 1, 1, 2, 2, 3
has_children.extend(Parent) (1, a), (2, b) each twice
Parent.aggr(colors, n="count(color)") (colors per parent) {1: 2, 2: 1, 3: 0} {1: 2, 2: 2, 3: 0}
Parent.aggr(has_children, n="count(*)", exclude_nonmatching=True) {1: 1, 2: 1} {1: 2, 2: 2}
has_children.aggr(Parent, n="count(label)") {1: 1, 2: 1} {1: 2, 2: 2}
dj.U("color").aggr(colors, n="count(*)") (parents per color) {blue: 2, red: 1} {blue: 3, red: 1}
dj.U().aggr(dj.U("color") & Child, n="count(*)") (number of colors) 2 4
len(Parent * has_children) 2 2 (correct, but to_dicts() returns 4 rows)
Parent & has_children, Parent - has_children parents 1, 2 / parent 3 correct

The SQL for Parent * has_children is a plain join with no DISTINCT:

SELECT `parent_id`,`label` FROM `parent` JOIN `child` USING (`parent_id`)

Cause

dj.U.__and__ (src/datajoint/expression.py:1507) returns a copy of the table expression with _distinct = True (line 1513). Only make_sql reads that flag (line 162), so DISTINCT appears only when the expression is the outermost query or is already wrapped in a subquery.

QueryExpression.join (line 343) decides whether to wrap each operand in a subquery (lines 400–425). The checks are:

  • the operands' source tables share columns other than the join columns
  • the operand is an aggregation or a union
  • a join column is computed
  • the operand already joins several tables

_distinct is not among them. An unwrapped operand is merged in by its tables, joins and conditions (lines 426–436). The result is a new expression with _distinct set to False, so the DISTINCT is lost. extend (line 490) and aggr (Aggregation.create, line 1341) go through the same join(), with the expression on either side.

When the operand is wrapped for another reason, the DISTINCT survives. For example, Child * (dj.U("parent_id") & Child) is correct, because both operands come from child and share child_id and color.

dj.U.aggr (line 1532) does not call join(). It builds the aggregation directly from the group's tables, joins and conditions (lines 1566–1573), and drops _distinct the same way.

__len__ (line 1149) counts distinct primary-key values (lines 1163–1174), which hides the repeated rows.

DataJoint 0.14 has the same code, so this is not new in 2.x.

Expected

An operand that has _distinct set should be wrapped in a subquery, as an aggregation or a union already is. That means wrapping it in join() for either operand, which covers *, extend, and aggr with the expression as the group or as the grouping side. dj.U.aggr needs the same wrapping for its group. Then SELECT DISTINCT … runs before the join or the GROUP BY. With this wrapping tried locally, every case in the table is correct on both backends, and the existing tests for dj.U, joins, aggr and grouping still pass.

#1657 adds the same kind of check for dj.Top at these places in join() and dj.U.aggr. It does not cover _distinct, and this reproduction fails the same way on that branch. The _distinct check belongs next to the _top one.

Ngôn ngữ chính
Python
Star
198
Fork
103
Merge trung bình
6 ngày 1 giờ
Pull request đã merge (30 ngày)
10

Chuẩn bị môi trường

Bắt đầu từ đâu

  1. Đọc hết issue, rồi đọc hướng dẫn đóng góp của dự án.
  2. Bình luận trên issue rằng bạn sẽ nhận — tránh hai người làm cùng một việc.
  3. Fork repository và làm thay đổi trên một nhánh.
  4. Mở pull request có tham chiếu số hiệu của issue.

Issue khác của datajoint/datajoint-python

Tất cả issue của datajoint/datajoint-python

Issue tương tự

Thêm issue về Python

Nhận issue mới trong hộp thư của bạn

Bản tóm tắt ngắn những issue GitHub phù hợp với người mới.