Practice data engineering interview problems from the terminal, in your editor: SQL against DuckDB, the same problems with DataFrames (pandas or polars), and plain Python problems (parsing logs, flattening JSON, ordering a DAG...).
uv tool install git+https://github.com/azure-eller/sqlgym
sqlgym # pick a mode and a level, then practice
sqlgym open 107 # open a specific problem (--df: with DataFrames)
sqlgym list # all problems and your progress
sqlgym check FILE # check a file's answer from the shellsqlgym starts with a menu: j/k (or ↑/↓) move between Mode and Level, h/l
(or ←/→) change them, Enter starts. Then it opens Neovim on two pages: the problem with your SQL below it, and
your SQL with the results below it. Running a query shows the results; the SQL
pane fits your query and the page gets the rest of the screen.
| Key | |
|---|---|
<leader>r |
run your query and check it |
<leader>n |
next problem (skips this one if unsolved) |
<leader>a |
show the answer |
<leader>R |
redo: clear your SQL (u undoes) |
Tab |
flip between the problem and the results |
Esc / q |
back to the problem from the results |
? |
show these keys |
Ctrl-j / Ctrl-k |
scroll the problem or results down / up |
:q |
save and quit |
The same actions are available as :Sqlgym run|next|answer|redo. The keys use
your own leader and exist only in sqlgym's buffers, so your Neovim config is
never changed. In those buffers sqlgym also turns off format-on-save,
completion popups and sqlfluff linting, which get in the way of practice.
Only the last statement in your SQL is checked, so you can keep scratch
queries above your answer. Wide tables and results scroll sideways (zl / zh).
Pick DataFrames in the menu (or sqlgym open 107 --df) to answer the SQL
problems in Python. Each problem starts you off with
import polars as pl # or: import pandas as pd
def solve(users):
returnsolve() gets the tables it names as DataFrames: polars, or pandas if your
file imports only pandas. Return a DataFrame, a Series or a single value. It's
checked against the SQL answer, so the same rules apply: columns by position,
floats to 4 places, and pandas' index is ignored unless it's named (as after a
groupby, where it becomes a column). print() output shows above your result.
Python problems (201+) are plain Python, as in the coding round of many DE
interviews: you write solve() with the arguments the problem names, and it's
run on several cases, each checked against the reference solution's result.
Dicts compare in any key order, lists and tuples alike, floats to 6 places.
A wrong answer shows the first case you got wrong, with what was expected.
Your code gets 10 seconds, so an endless loop just fails.
Progress lives in ~/.local/share/sqlgym/progress.json and your answers in
~/.local/share/sqlgym/work/<NNN_slug>.sql (or .py); both respect $XDG_DATA_HOME.
problems/<NNN_slug>/
problem.toml # title, difficulty, ordered (default false)
prompt.md # the question text
setup.sql # CREATE TABLE + INSERT statements
solution.sql # reference answer
solution.py # optional: the DataFrame answer `<leader>a` shows
A Python problem has problem.toml and prompt.md, then cases.py (a list of
keyword arguments, cases = [dict(logs=[...]), ...]) and solution.py (a
solve()) instead of the SQL files. The first case is shown as the example.
The schema and sample rows shown with each problem are generated from setup.sql.
Run uv run pytest to check that every solution runs (and passes its cases).
Problems 001–099 come from pgexercises
and are regenerated by uv run scripts/import_pgexercises.py, which writes
scripts/import_report.md. Problems 101+ are original.
The code is MIT. Problems 001–099 are adapted from pgexercises by Alisdair Owens and stay under CC BY-SA 3.0. See LICENSE and problems/LICENSE.
