If you care about performance, you may want to avoid CSV files. But since our data sources are often like our family, we can't make a choice, we'll see in this blog post how to process a CSV file as fast as possible.

top 7 comments

sorted by: hot top controversial new old

[–] [email protected] 27 points 8 months ago (1 children)

Holy shit, switching to PyArrow is going to make me seem a mystical wizard when I merge in the morning. I’ve easily halved the execution time of a horrible but unavoidable job (yay crappy vendor “API” that returns a huge CSV).

[–] [email protected] 3 points 8 months ago

You and me both. I've been parsing around 10-100 million row CSVs lately and...this will hopefully help.

[–] [email protected] 7 points 8 months ago

LOL just use fscanf() you silly goose

[–] [email protected] 4 points 8 months ago* (last edited 8 months ago) (3 children)

Okay so would it be faster to convert it to something better and then do something faster with this better format?

Edit: I guess looking at the numbers, they’re already pretty low there. Idk how much faster it’d really be and whether not it’d be worth doing

What’s even the “gold standard” for logging stuff I guess?

[–] NostraDavid 6 points 8 months ago

What’s even the “gold standard” for logging stuff I guess?

structlog. Or just Structured Logging in general.

Don't do:

logging.info(f"{something} happened!")

But do

logging.info("thing-happened", thing=something)

Why? Your event will become a category, which means it's easily searchable/findable, you can output either human-readable stuff (the typical {date}, {loglevel}, {event}) or just straight up JSONL (a JSON object/dict per line). If you have JSON logs you can use jq to query/filter/manipulate your logs, if you have something like ELK, you can insert your logs there and create dashboards.

It's amazing - though it may break your brain initially.

[–] [email protected] 2 points 8 months ago

That really depends on how much of it you're doing. If you're just handing a few times at a time, the difference between 0.1s and 3s isn't that big of a deal. If you're handling thousands or even millions in a day, it can be an order of magnitude cost savings to make it more efficient.

We use a CSVs at work, but it's not a common thing so we just use the built-in csv library. If we did more with it, pandas would be the way to go (or maybe we'd rewrite that service in Rust).

[–] NostraDavid 2 points 8 months ago

Also, regarding better formats: parquet is relatively nice. Smaller files, though not human readable. Use parquet if you read often, or have IO issues (file "too large" as CSV).