Join Order Benchmark

General This is a re-distribution of the IMDB dataset, adapted for different database systems. The dumps was obtained from the official IMDB servers at a later time than the original publication. Therefore, it contains significantly more data. For example, the title relation contains approx. 4.7M tuples instead of 2.5M. The dumps contain all possible foreign key constraints. In addition to the IMDB dataset, we also provide different variants of the Join Order Benchmark (JOB): the original JOB workload as introduced by Leis et al. in their paper "How Good Are Query Optimizers, Really?" (Proc. VLDB ,Vol. 9, Issue 3) (DOI 10.14778/2850583.2850594) the simplified JOB-light version as introduced by Kipf et al. in their paper "Learned Cardinalities: Estimating Correlated Joins with Deep Learning" (CIDR 2019) (Link) the extended JOB-Complex workload as introduced by Wehrstein et al. in their paper "JOB-Complex: A Challenging Benchmark for Traditional and Learned Query Optimization" (AIDB 2025) (Link) Import for PostgreSQL The PostgreSQL dump was created using PostgreSQL 16, so it should work on all Postgres version starting from PG 16. To load the dump, first create a new Postgres database. Afterwards, run pg_restore on the dump: $ createdb imdb $ pg_restore -n public -O -d imdb imdb_dump.pg If you do not have the pg_hint_plan extension available on your server, you might see a couple of warnings during the import. These are safe to ignore. For reference, we also provide the raw schema for Postgres. Similarly, a mismatch of the user name may also produce some error messages. You can ignore those to, Postgres proceeds with the restore anyway. Import for DuckDB Simply run DuckDB on the dump, no further steps required. $ duckdb imdb.duckdb For reference, we also provide the raw schema for DuckDB.