exp show: using wrong results for display

DVC version: 2.1.0+fc4e31 (osxpkg)
---------------------------------
Platform: Python 3.8.2 on macOS-11.3.1-x86_64-i386-64bit
Supports: All remotes
Cache types: reflink, hardlink, symlink
Cache directory: apfs on /dev/disk1s1s1
Caches: local
Remotes: https
Workspace directory: apfs on /dev/disk1s1s1
Repo: dvc, git

Seems like #5880 introduced some bug: Given experiments runs:

def test_something(tmp_dir, scm, dvc, exp_stage):
    baseline_rev = scm.get_rev()
    exp1 = dvc.experiments.run(exp_stage.addressing, params=["foo=2"])
    exp2 = dvc.experiments.run(exp_stage.addressing, params=["foo=3"])

    result = dvc.experiments.show()
    rev1 = first(exp1)
    rev2 = first(exp2)

    assert result[baseline_rev][rev1]["params"]["params.yaml"] == {"foo": 2}
    assert result[baseline_rev][rev2]["params"]["params.yaml"] == {"foo": 3}

Note that the asserts seem to verify that internals work correctly if we go to the test directory and run dvc exp show, we will get:

┏━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━┳━━━━━┓
┃ Experiment    ┃ Created  ┃ foo ┃ foo ┃
┡━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━╇━━━━━┩
│ workspace     │ -        │   3 │ 3   │
│ master        │ 04:54 PM │   3 │ 3   │
│ ├── exp-e0d04 │ 04:54 PM │   1 │ 1   │
│ └── exp-42ced │ 04:54 PM │   1 │ 1   │
└───────────────┴──────────┴─────┴─────┘

while it should more like:

┏━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━┳━━━━━┓
┃ Experiment    ┃ Created  ┃ foo ┃ foo ┃
┡━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━╇━━━━━┩
│ workspace     │ -        │   3 │ 3   │
│ master        │ 04:54 PM │   1 │ 1   │
│ ├── exp-42ced │ 04:54 PM │   2 │ 2   │
│ └── exp-e0d04 │ 04:54 PM │   3 │ 3   │
└───────────────┴──────────┴─────┴─────┘

cc @skshetry

[skshetry]

The issue here is related to the columns having the same name. It’s quite interesting.

So, what we do in the table is that we keep each column in a buffer/list. So, when we are adding a row of data here, it’s pushing data twice for those columns, and at the end, we get a buffer overrun beyond it’s position.

So, when you are adding rows [(3,3), (1, 1), (2, 2), (3,3)], it is being added to the foo column as [3, 3, 1, 1, 2, 2, 3, 3]. And you’ll get a truncated version [3, 3, 1, 1, 2, 2] during the rendering. And, since it’s the same column, you get the same data in both of the columns.

There are two issues:

1. Garbled/garbage data in the columns.
2. Styling of the column is affected.

The easiest way might be to not normalize headers for metrics (or take both into account). Alternatively, we could also fix this by adding support for the same column names, but then we’ll need to address them by index.

[pared]

The problem in this issue is that we are seeing foo=1 for both experiments, while for exp-42ced it should be 2 and for exp-e0d04 it should be 3