GC string views on hash join build side

[ctsk]

Which issue does this PR close?

• Closes Excessive Arc-clone in HashJoinStream with StringView on build-side #16206.

What changes are included in this PR?

• A utility function to garbage collect (gc) all view-type columns of a batch.
• Perform gc on the build-side batch of a hash join

Are these changes tested?

Are there any user-facing changes?

No.

[ctsk]

Benchmark results:

Comparing main and fix_build-side-gc
--------------------
Benchmark tpch_sf10.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query        ┃       main ┃ fix_build-side-gc ┃        Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 1     │ 2273.99 ms │        2561.92 ms │  1.13x slower │
│ QQuery 2     │  414.94 ms │         420.20 ms │     no change │
│ QQuery 3     │  978.78 ms │         942.35 ms │     no change │
│ QQuery 4     │  424.50 ms │         479.00 ms │  1.13x slower │
│ QQuery 5     │ 1206.94 ms │        1402.47 ms │  1.16x slower │
│ QQuery 6     │  455.31 ms │         222.48 ms │ +2.05x faster │
│ QQuery 7     │ 1803.05 ms │        1657.75 ms │ +1.09x faster │
│ QQuery 8     │ 1069.79 ms │        1520.80 ms │  1.42x slower │
│ QQuery 9     │ 2336.36 ms │        2277.62 ms │     no change │
│ QQuery 10    │ 1061.73 ms │         948.20 ms │ +1.12x faster │
│ QQuery 11    │  288.41 ms │         294.14 ms │     no change │
│ QQuery 12    │  996.93 ms │         726.61 ms │ +1.37x faster │
│ QQuery 13    │  692.57 ms │         832.22 ms │  1.20x slower │
│ QQuery 14    │  315.32 ms │         489.95 ms │  1.55x slower │
│ QQuery 15    │  814.54 ms │         694.10 ms │ +1.17x faster │
│ QQuery 16    │  299.17 ms │         223.10 ms │ +1.34x faster │
│ QQuery 17    │ 2000.39 ms │        1892.36 ms │ +1.06x faster │
│ QQuery 18    │ 2443.55 ms │        2400.40 ms │     no change │
│ QQuery 19    │  532.30 ms │         796.33 ms │  1.50x slower │
│ QQuery 20    │  904.87 ms │         759.41 ms │ +1.19x faster │
│ QQuery 21    │ 2586.84 ms │        2944.05 ms │  1.14x slower │
│ QQuery 22    │  456.56 ms │         414.87 ms │ +1.10x faster │
└──────────────┴────────────┴───────────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Benchmark Summary                ┃            ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ Total Time (main)                │ 24356.86ms │
│ Total Time (fix_build-side-gc)   │ 24900.32ms │
│ Average Time (main)              │  1107.13ms │
│ Average Time (fix_build-side-gc) │  1131.83ms │
│ Queries Faster                   │          9 │
│ Queries Slower                   │          8 │
│ Queries with No Change           │          5 │
│ Queries with Failure             │          0 │
└──────────────────────────────────┴────────────┘
--------------------
Benchmark tpch_sf1.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query        ┃      main ┃ fix_build-side-gc ┃        Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 1     │ 349.95 ms │         355.99 ms │     no change │
│ QQuery 2     │  74.36 ms │          43.87 ms │ +1.69x faster │
│ QQuery 3     │  55.24 ms │          59.94 ms │  1.09x slower │
│ QQuery 4     │  46.10 ms │          71.84 ms │  1.56x slower │
│ QQuery 5     │ 122.83 ms │         147.10 ms │  1.20x slower │
│ QQuery 6     │  60.77 ms │          30.01 ms │ +2.02x faster │
│ QQuery 7     │ 151.49 ms │         234.96 ms │  1.55x slower │
│ QQuery 8     │  79.57 ms │          99.71 ms │  1.25x slower │
│ QQuery 9     │ 210.99 ms │         202.58 ms │     no change │
│ QQuery 10    │ 132.90 ms │         144.02 ms │  1.08x slower │
│ QQuery 11    │  38.18 ms │          33.87 ms │ +1.13x faster │
│ QQuery 12    │ 118.61 ms │         167.74 ms │  1.41x slower │
│ QQuery 13    │  55.91 ms │          60.70 ms │  1.09x slower │
│ QQuery 14    │  76.92 ms │          62.10 ms │ +1.24x faster │
│ QQuery 15    │  84.87 ms │          60.17 ms │ +1.41x faster │
│ QQuery 16    │  23.91 ms │          25.35 ms │  1.06x slower │
│ QQuery 17    │ 145.09 ms │         182.89 ms │  1.26x slower │
│ QQuery 18    │ 228.53 ms │         244.52 ms │  1.07x slower │
│ QQuery 19    │  96.24 ms │          66.66 ms │ +1.44x faster │
│ QQuery 20    │  56.56 ms │         104.25 ms │  1.84x slower │
│ QQuery 21    │ 199.91 ms │         230.70 ms │  1.15x slower │
│ QQuery 22    │  45.35 ms │          49.77 ms │  1.10x slower │
└──────────────┴───────────┴───────────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
┃ Benchmark Summary                ┃           ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
│ Total Time (main)                │ 2454.30ms │
│ Total Time (fix_build-side-gc)   │ 2678.74ms │
│ Average Time (main)              │  111.56ms │
│ Average Time (fix_build-side-gc) │  121.76ms │
│ Queries Faster                   │         6 │
│ Queries Slower                   │        14 │
│ Queries with No Change           │         2 │
│ Queries with Failure             │         0 │
└──────────────────────────────────┴───────────┘

Seems like it slows down SF 1, 10 a bit :/

[ctsk]

At SF=100, this PR is 10% faster:

--------------------
Benchmark tpch_sf100.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query        ┃         pr ┃         old ┃        Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 1     │ 1728.00 ms │  1755.28 ms │     no change │
│ QQuery 2     │ 1356.73 ms │  1241.24 ms │ +1.09x faster │
│ QQuery 3     │ 1987.91 ms │  1995.85 ms │     no change │
│ QQuery 4     │ 1509.51 ms │  1593.78 ms │  1.06x slower │
│ QQuery 5     │ 3676.40 ms │  3770.53 ms │     no change │
│ QQuery 6     │  277.50 ms │   287.08 ms │     no change │
│ QQuery 7     │ 3542.47 ms │  3499.00 ms │     no change │
│ QQuery 8     │ 4037.40 ms │  4228.35 ms │     no change │
│ QQuery 9     │ 5968.77 ms │  5910.46 ms │     no change │
│ QQuery 10    │ 1716.65 ms │  1666.84 ms │     no change │
│ QQuery 11    │ 1059.97 ms │  1005.02 ms │ +1.05x faster │
│ QQuery 12    │ 1091.32 ms │  1110.42 ms │     no change │
│ QQuery 13    │ 1121.01 ms │  1234.91 ms │  1.10x slower │
│ QQuery 14    │  587.82 ms │   535.24 ms │ +1.10x faster │
│ QQuery 15    │  904.97 ms │   890.88 ms │     no change │
│ QQuery 16    │  731.39 ms │   745.86 ms │     no change │
│ QQuery 17    │ 5760.02 ms │  5919.10 ms │     no change │
│ QQuery 18    │ 6404.64 ms │  8073.58 ms │  1.26x slower │
│ QQuery 19    │  688.99 ms │   666.94 ms │     no change │
│ QQuery 20    │ 1725.25 ms │  1742.06 ms │     no change │
│ QQuery 21    │ 7427.94 ms │ 10795.91 ms │  1.45x slower │
│ QQuery 22    │ 1373.74 ms │  1370.31 ms │     no change │
└──────────────┴────────────┴─────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Benchmark Summary      ┃            ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ Total Time (fx)        │ 54678.41ms │
│ Total Time (HEAD)      │ 60038.64ms │
│ Average Time (fx)      │  2485.38ms │
│ Average Time (HEAD)    │  2729.03ms │
│ Queries Faster         │          3 │
│ Queries Slower         │          4 │
│ Queries with No Change │         15 │
│ Queries with Failure   │          0 │
└────────────────────────┴────────────┘

Perhaps a threshold is needed?

[2010YOUY01]

Thank you! This is great. I got some minor suggestions:

1. The issue explained the motivation of the change clearly, I recommend to add the same rationale to the code comment to make the code more understandable.
2. We can unify gc_record_batch() and this existing function

   datafusion/datafusion/physical-plan/src/sorts/sort.rs

   Line 469 in 3c4e39a

   fn organize_stringview_arrays(

   I'm also happy to take this on as a follow-up task.

[ctsk]

I will do both of these things later today. I am concerned about the performance impact for smaller-scale tasks. I suspect many users of datafusion are not doing such large joins....

[Dandandan]

We have quite some implementations of gc-ing arrays. I am wondering in this case if the performance can be improved for smaller tables by this heuristic used here:

https://github.com/apache/datafusion/blob/3c4e39ac0cf83bd8ead45722a5873bac731b53f1/datafusion/physical-plan/src/coalesce/mod.rs#L222C4-L222C24