| --- |
| license: mit |
| language: |
| - en |
| pretty_name: key-order |
| size_categories: |
| - n<1K |
| tags: |
| - postgres |
| - database |
| - benchmark |
| - uuid |
| - btree |
| configs: |
| - config_name: measurements |
| data_files: |
| - split: train |
| path: out/measurements-pass*.jsonl |
| - config_name: reads |
| data_files: |
| - split: train |
| path: out/reads-pass*.jsonl |
| - config_name: measurements_docker |
| data_files: |
| - split: train |
| path: out-docker/measurements-pass*.jsonl |
| --- |
| |
| # key-order |
|
|
| Author: Dmitriy Semenkevich (dimhold) |
|
|
| This dataset is an independent replication of the cost of a random UUID primary key against ordered keys on Postgres: 3 key strategies (random uuid, uuid v7, bigint TSID) inserted into identical tables, with every batch timing and buffer counter. |
|
|
| Headline, from the repository: after three million rows the index on random keys has 49.8% leaf fragmentation and 66.7% leaf density, while the index on time ordered keys has 0% fragmentation and 90% density, on identical data. Reads cost the same: a point lookup answered in 236 us cold on both indexes. |
|
|
| This is a mirror. The canonical home is the Zenodo record and the GitHub repository; criteria, limitations and full tables live there. |
|
|
| - Paper: https://dimhold.by/papers/key-order/ |
| - Code and full write up: https://github.com/dimhold/key-order |
| - DOI: https://doi.org/10.5281/zenodo.22128828 |
|
|
| ## Files |
|
|
| | path | what it is | |
| |---|---| |
| | `out/measurements-pass*.jsonl` | insert passes: per strategy sizes, `pgstatindex`, batch timings, settings | |
| | `out/reads-pass*.jsonl` | cold and warm point lookups with buffer counters | |
| | `out-docker/measurements-pass*.jsonl` | the same insert passes in Docker | |
|
|
| Files ending in `.jsonl` here are `.ndjson` in the GitHub repository, renamed so the viewer reads them; the content is unchanged. |
|
|
| ## How it was produced |
|
|
| Keys are generated deterministically ahead of time into a source table identical for every strategy, `fsync` stays on and the strategy order alternates between passes. The disproof threshold was fixed before counting (`CRITERIA.md` in the repository). No model was involved. The scripts that turn these files into every number are in the GitHub repository. |
|
|
| ## Licence |
|
|
| MIT, the licence of the repository, covering code and data. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{semenkevich2026keyorder, |
| author = {Semenkevich, Dmitriy}, |
| title = {key-order: what a random primary key costs on Postgres, an independent replication}, |
| year = {2026}, |
| doi = {10.5281/zenodo.22128828}, |
| url = {https://doi.org/10.5281/zenodo.22128828} |
| } |
| ``` |
|
|
