Pair programming with ChatGPT

A few days ago, ChatGPT was released for the entire world to marvel and wonder whether our Adam-inherited struggles through life had come to an end, thanks to our eternal eagerness to fly higher than Icarus, past the sun, reaching for the stars.

ChatGPT is a general-purpose chatbot that you can simply talk to using your keyboard as a human-to-machine interface. I ran a few experiments, and I can tell you this is the closest thing I could ever imagine to what is described in Her.

Now let us jump to recent research that I've been running for the last couple of days. As a data engineer, I struggle to find a high-performance solution to automatically create time-based table partitions in PostgreSQL at data insertion time.

When a new record is appended to a table, the relative partition should be created on-the-fly if it didn't exist already.

I think I found a decent solution to this issue, but I'm not going to bother you with that in this article. What is relevant here is that I ran the question to ChatGPT:

How do I create automated time-based partitions in Postgres?

This was the starting point of an experience that I can only refer to as a pair programming session in which I held the Navigator role.

The bot began issuing explanations and code examples, all relevant material but rather incomplete. It looked like it was simply rewriting (for the better) the standard PostgreSQL documentation about partitioning.

I can't tell you why, but it chose to partition by year, and I wasn't satisfied by it, so I gave my feedback:

I would like the partition to be by week

This is just plain English, with a contextual reference to the results that it had just provided. It f****g worked. I got a new code with the correct time rules applied to it.

Plus, here are some excellent explanations regarding INTERVALS in PostgreSQL!

I can continue with the detailed report, but my wife wants me to run to the supermarket, buy lasagna, and cook our Christmas meal. So I let you with the full recording of the session:

The end of the story is... we (programmers) are screwed. What it once took days to discover now takes seconds.

I ran a few other experiments, and I got the following:

• A correct-ish implementation of the PushID algorithm in plpgsql, though I had to give a lot of guidance
• A correct docker-compose.yml to run ElasticSearch with minimal memory footprint – correlated by the explanations why that is a bad idea
• A correct explanation of an APPEND-ONLY data structure in PostgreSQL – with a content quality so high that I can easily copy/paste it and create a new article out of it

I used to be the hero, but now I feel like I will be the sidekick.