Use compute.FilterRecord instead of bitmask and manually filtering about frostdb

[gernest]

I took another look, we already have arrowutils.ReorderRecord which takes care of all the case I mentioned. We can go back to @thorfour sample and build array.Int32 instead of *array.Boolean from the bitmap. Pretty much calling (*Bitmap).ToArray() will give you the indices or we can avoid allocation and iterate on the bitmap to build the indices array.

Something like

	b := array.NewInt32Builder(...)
	b.Reserve(int(bitmap.GetCardinality()))
	bitmap.Iterate(func(x uint32) bool {
		b.Append(int32(x))
		return true
	})
	indices := b.NewArray()

So steps will be

• *Bitmap -> *array.Int32
• call arrowutils.ReorderRecord with the result
• Profit

[thorfour]

Sorry, apparently fat fingered the close button.

I played around with this today

 func filter(ctx context.Context, pool memory.Allocator, filterExpr BooleanExpression, ar arrow.Record) (arrow.Record, bool, error) {
      bitmap, err := filterExpr.Eval(ar)
      if err != nil {
          return nil, true, err
      }

      if bitmap.IsEmpty() {
          return nil, true, nil
      }

      // Construct filter array
      // NOTE: this is intermediary right now. The Eval function should return a boolean array instead so we can directly pass it in.
      bldr := array.NewBooleanBuilder(pool)
      defer bldr.Release()
      for i := 0; i < int(ar.NumRows()); i++ {
          bldr.Append(bitmap.Contains(uint32(i)))
      }
      filterArr := bldr.NewArray()
      defer filterArr.Release()

      result, err := compute.FilterRecordBatch(ctx, ar, filterArr, compute.DefaultFilterOptions())
      if err != nil {
          return nil, true, err
      }
      return result, false, nil
  }

And ended up with a panic

--- FAIL: Test_DB_All (0.00s)
    db_test.go:2085:
        	Error Trace:	/Users/thor/go/src/github.com/polarsignals/frostdb/db_test.go:2085
        	Error:      	Received unexpected error:
        	            	not implemented: function 'array_take' has no kernel matching input types (dictionary<values=utf8, indices=uint32, ordered=false>, uint16)
        	Test:       	Test_DB_All

Looks like it's not implemented for all types.

[garrensmith]

That is unfortunate. Should we open an issue on apache arrow for this?

[thorfour]

Yea and link it here so we know if/when we can actually implement this.

[gernest]

Yea and link it here so we know if/when we can actually implement this.

No need for this. We can in fact implement it now. Basically, this is how filter on records works

• build array of indices containing rows you wan't to choose
• for each record column use array_take to select rows of interest
• assemble result of previous step (concurrently)
• build a new record with the result.

So, actually @thorfour solution is correct, we just need to be smart with array.Dictionary columns. We can be smart, and take and assemble ourselves when we know we have *array.Dictionary column and delegate to compute for the rest.

So, @thorfour can you please open the PR with your changes ? I will help and make sure we massage it until it works for any case we currently have, we will probably need to run benchmarks as well to make sure we don't introduce regressions.

[gernest]

Quick check says it is not a simple change. I'm taking this task. Will submit supplementary patches to make it possible.

[gernest]

😓 finally I have this working. I will wait for supplementary patches to land then I will drop the PR.

[gernest]

Done #697 . I wish someone with access to production like workload would do some benchmarks and give us numbers.

[github-actions]

This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.