colserde: add simple serialization for DECIMALs

[yuzefovich]

This commit adds simple serialization of DECIMALs using the provided
MarshalText method.

Touches: #37044.

Release note (sql change): Previously, DECIMALs could not be sent over
the network when the computation was performed by the vectorized engine.
This has been fixed, and the vectorized engine now fully supports
DECIMAL type.

[cockroach-teamcity]

This change is 

[yuzefovich]

Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @asubiotto)

---

pkg/col/colserde/arrowbatchconverter_test.go, line 102 at r1 (raw file):

				}
			}
		} else if typ == coltypes.Timestamp {

I've just run into a failure because require.Equal didn't work on equivalent timestamp slices. This probably happened because there is *time.Location in there, and maybe my assumption that those locations are cached (which I mentioned in pkg/sql/colexec/allocator.go:172) is incorrect. But it's a separate concern, so I'll leave it as TODO.

[yuzefovich]

The benchmarks look rather sad though:

BenchmarkArrowBatchConverter/Bool/nullFraction=0.00/BatchToArrow-16         	  447144	      2249 ns/op	 455.37 MB/s
BenchmarkArrowBatchConverter/Bool/nullFraction=0.25/BatchToArrow-16         	  542944	      2255 ns/op	 454.18 MB/s
BenchmarkArrowBatchConverter/Bool/nullFraction=0.50/BatchToArrow-16         	  546918	      2258 ns/op	 453.57 MB/s
BenchmarkArrowBatchConverter/Bool/nullFraction=0.00/ArrowToBatch-16         	  620725	      1875 ns/op	 546.18 MB/s
BenchmarkArrowBatchConverter/Bool/nullFraction=0.25/ArrowToBatch-16         	  672000	      1829 ns/op	 559.94 MB/s
BenchmarkArrowBatchConverter/Bool/nullFraction=0.50/ArrowToBatch-16         	  666283	      1827 ns/op	 560.52 MB/s
BenchmarkArrowBatchConverter/Bytes/nullFraction=0.00/BatchToArrow-16        	 5882042	       200 ns/op	327924.15 MB/s
BenchmarkArrowBatchConverter/Bytes/nullFraction=0.25/BatchToArrow-16        	 5831419	       205 ns/op	320126.79 MB/s
BenchmarkArrowBatchConverter/Bytes/nullFraction=0.50/BatchToArrow-16        	 5823836	       204 ns/op	320532.64 MB/s
BenchmarkArrowBatchConverter/Bytes/nullFraction=0.00/ArrowToBatch-16        	  769612	      1558 ns/op	42070.03 MB/s
BenchmarkArrowBatchConverter/Bytes/nullFraction=0.25/ArrowToBatch-16        	  748358	      1628 ns/op	40267.03 MB/s
BenchmarkArrowBatchConverter/Bytes/nullFraction=0.50/ArrowToBatch-16        	  674947	      1619 ns/op	40490.85 MB/s
BenchmarkArrowBatchConverter/Decimal/nullFraction=0.00/BatchToArrow-16      	    3733	    319850 ns/op	  71.26 MB/s
BenchmarkArrowBatchConverter/Decimal/nullFraction=0.25/BatchToArrow-16      	    3742	    319979 ns/op	  71.23 MB/s
BenchmarkArrowBatchConverter/Decimal/nullFraction=0.50/BatchToArrow-16      	    3673	    319458 ns/op	  71.35 MB/s
BenchmarkArrowBatchConverter/Decimal/nullFraction=0.00/ArrowToBatch-16      	    2220	    538115 ns/op	  42.36 MB/s
BenchmarkArrowBatchConverter/Decimal/nullFraction=0.25/ArrowToBatch-16      	    2264	    530742 ns/op	  42.95 MB/s
BenchmarkArrowBatchConverter/Decimal/nullFraction=0.50/ArrowToBatch-16      	    2154	    534273 ns/op	  42.66 MB/s
BenchmarkArrowBatchConverter/Int64/nullFraction=0.00/BatchToArrow-16        	 7542470	       157 ns/op	52246.90 MB/s
BenchmarkArrowBatchConverter/Int64/nullFraction=0.25/BatchToArrow-16        	 7385175	       163 ns/op	50367.65 MB/s
BenchmarkArrowBatchConverter/Int64/nullFraction=0.50/BatchToArrow-16        	 7430563	       161 ns/op	50768.17 MB/s
BenchmarkArrowBatchConverter/Int64/nullFraction=0.00/ArrowToBatch-16        	 8732474	       136 ns/op	60347.13 MB/s
BenchmarkArrowBatchConverter/Int64/nullFraction=0.25/ArrowToBatch-16        	 6350826	       189 ns/op	43292.38 MB/s
BenchmarkArrowBatchConverter/Int64/nullFraction=0.50/ArrowToBatch-16        	 6361404	       190 ns/op	43080.26 MB/s
BenchmarkArrowBatchConverter/Timestamp/nullFraction=0.00/BatchToArrow-16    	   15110	     80315 ns/op	 306.00 MB/s
BenchmarkArrowBatchConverter/Timestamp/nullFraction=0.25/BatchToArrow-16    	   14852	     80079 ns/op	 306.90 MB/s
BenchmarkArrowBatchConverter/Timestamp/nullFraction=0.50/BatchToArrow-16    	   14872	     80608 ns/op	 304.88 MB/s
BenchmarkArrowBatchConverter/Timestamp/nullFraction=0.00/ArrowToBatch-16    	    9068	    132900 ns/op	 184.92 MB/s
BenchmarkArrowBatchConverter/Timestamp/nullFraction=0.25/ArrowToBatch-16    	    9212	    132726 ns/op	 185.16 MB/s
BenchmarkArrowBatchConverter/Timestamp/nullFraction=0.50/ArrowToBatch-16    	    9378	    132135 ns/op	 185.99 MB/s

[asubiotto]

Thanks for this! Could you try running some higher-level benchmarks to see what the microbenchmark slowdown translates to when running a query? Maybe some distributed aggregation on decimals just above the row count threshold to get a worst case comparison.

[yuzefovich]

I ran a simple benchmark as you suggested: I set up a three node local cluster and created a table with 1000 decimals which were split into 3 ranges with one range per node and then ran SELECT sum(d) FROM t. With vectorized engine the query took around 1.5 ms, with row-by-row engine 1.2 ms. Then inserted another 9000 decimals and ran the query again. Now vectorized took 2.9 ms whereas row-by-row 3.2 ms. I think the performance of the current serialization is more or less acceptable, but if we have time this release to improve it, we should do so. In any case, it is (probably in most cases) better than not having serialization at all.

[asubiotto]

Reviewed 5 of 6 files at r1, 2 of 2 files at r2.
Reviewable status: complete! 1 of 0 LGTMs obtained

[yuzefovich]

TFTR!

bors r+

[craig]

Build succeeded

• GitHub CI (Cockroach)