In testing, LDBC 100 loads in 3 seconds on ac4 (with 128 cores). However, this does not account for the overhead of counting the number of rows. If we are not going to make hash indexes resizable soon, we should add a dedicated CSV row counting function.

Furthermore, the 3 seconds does not account for the overhead of writing to disk, since I ran:

LOAD FROM <csv_path> RETURN COUNT(*)

On LDBC-10, which is much nicer to benchmark because the serial CSV reader loads it quickly enough, I got these numbers:

Serial: ~12s.
Parallel: ~350ms.

Again, in practice, we see only a 2x speedup since we pay the price of counting the rows.

Codecov Report

Attention: 6 lines in your changes are missing coverage. Please review.

Comparison is base (f05d348) 89.55% compared to head (495e283) 89.60%.
Report is 4 commits behind head on master.

@@            Coverage Diff             @@
##           master    #2070      +/-   ##
==========================================
+ Coverage   89.55%   89.60%   +0.04%     
==========================================
  Files         981      985       +4     
  Lines       35901    35745     -156     
==========================================
- Hits        32151    32028     -123     
+ Misses       3750     3717      -33     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.