Medium

Nazmul Hasan

Featured Book

See all (2)

Shipping AI-Written Code: Review, Verification and Guardrails for Teams Running Coding Agents

Shipping AI-Written Code: Review, Verification and Guardrails for Teams Running Coding Agents

A practical guide for engineering teams to review, verify and add guardrails to code written by AI coding agents, fast enough to keep up.

2026

Stories

Pinned

In

Artificial Intelligence in Plain English

by

Nazmul Hasan

·

Aug 26

I Measured Every Skill Anthropic Ships. Five of Them Break the Budget.

Progressive disclosure explains how Agent Skills load. It says nothing about what they cost once they are in context, and a scan of the…

I Measured Every Skill Anthropic Ships. Five of Them Break the Budget.

Pinned

In

Data Science Collective

by

Nazmul Hasan

·

Sep 9

DuckDB vs Postgres vs SQLite: The Same Query Takes Postgres 80 Seconds and DuckDB 0.32.

The crossover everyone asks about is not a row count. It is one query shape, and it breaks at a million rows.

DuckDB vs Postgres vs SQLite: The Same Query Takes Postgres 80 Seconds and DuckDB 0.32.

Pinned

In

Data Science Collective

by

Nazmul Hasan

·

Aug 31

One Argument Makes pandas.read_csv 11.6× Faster. Most People Migrate Instead.

Seven readers at 50,000, 1 million and 8 million rows — and the fixed cost that reverses the ranking on small files.

Conceptual illustration of a single striped file feeding channels of very different widths.

Pinned

In

Data Science Collective

by

Nazmul Hasan

·

Aug 29

I Wrote the Same 25 Million Rows 17 Ways. The Fastest Layout Was 57× Quicker Than the Slowest.

Row group size, codec, dictionary encoding and sort order, measured — plus the one-line query mistake that undoes all of it.

I Wrote the Same 25 Million Rows 17 Ways. The Fastest Layout Was 57× Quicker Than the Slowest.

Pinned

In

Data Science Collective

by

Nazmul Hasan

·

Jun 4

3 Costly Mistakes You Are Making With Pandas (And How to Fix Them With Polars)

Your data pipeline isn’t slow because of your laptop. It’s slow because of habits you picked up in 2015.

3 Costly Mistakes You Are Making With Pandas (And How to Fix Them With Polars)

Nazmul Hasan

·

22h ago

Postgres Connection Pooling in Python, Measured: Reconnecting Never Passed 1,000 req/s

psycopg_pool served 14,000 to 20,000 on the same laptop. I also tested SQLAlchemy’s pool, PgBouncer, and the setting that let psycopg 3…

Postgres Connection Pooling in Python, Measured: Reconnecting Never Passed 1,000 req/s

Nazmul Hasan

·

1d ago

Postgres Partitioning Didn’t Speed Up Reads, but Dropping Old Data Got 140x Faster

20M rows on PostgreSQL 18.6: range reads tied a plain b-tree, lookups by id got 3x to 700x slower, and a file limit explained most of the…

Postgres Partitioning Didn’t Speed Up Reads, but Dropping Old Data Got 140x Faster

Nazmul Hasan

·

2d ago

pandas read_sql Needed 1.1 GB to Read a 98 MB Postgres Table

I timed 15 ways to read Postgres into pandas and Polars. ConnectorX and DuckDB were 9 to 12x faster on a quarter of the memory.

pandas read_sql Needed 1.1 GB to Read a 98 MB Postgres Table

Nazmul Hasan

·

3d ago

dbt Incremental Strategies Compared: Every One Left Wrong Rows in My Test

Tested on DuckDB and Postgres from 100k to 32M rows. A one-day lookback on merge fixed every wrong row; here is what each strategy costs as…

dbt Incremental Strategies Compared: Every One Left Wrong Rows in My Test

Nazmul Hasan

·

4d ago

Postgres OFFSET vs Keyset Pagination: Up to 9 Seconds vs 0.07 ms on 10M Rows

Keyset only stays flat with an id tie-breaker and a row-value WHERE clause. Here’s the Python.

Postgres OFFSET vs Keyset Pagination: Up to 9 Seconds vs 0.07 ms on 10M Rows

Nazmul Hasan

Nazmul Hasan

Book Author

I benchmark databases, Python tools, and AI coding agents using real workloads—then publish the numbers, code, and unexpected trade-offs.

Help

Status

About

Careers

Press

Blog

Store

Privacy

Rules

Terms

Text to speech