Hacker News Reader

Comment by vovavili | Hacker News Reader

Replacing an 11.6GB Parquet file every 5 minutes strikes me as a bit wasteful. I would probably use Apache Iceberg here.

View on HN

It's not doing that. If you look at the repository, it's adding a new commit with tiny parquet files every 5 minutes. This recent one only was a 20.9 KB parquet file: https://huggingface.co/datasets/open-index/hacker-news/commi... and the ones before it were a median of 5 KB: https://huggingface.co/datasets/open-index/hacker-news/tree/...

The bigger concern is how large the git history is going to get on the repository.

I recall that this became a big problem for the Homebrew project in terms of load on the repo, to the extent that Github asked them not to recommend/default-enable shallow clones for their users: https://github.com/Homebrew/brew/issues/15497#issuecomment-1...

This is likely to be lower traffic, and the history should (?) scale only linearly with new data, so likely not the worst thing. But it's something to be cognizant of when using SCM software in unexpected ways!

How would shallow clone be more stressful for GitHub than a regular clone?

Shallow clones (and the resulting lack of shared history data) break many assumptions that packfile optimisations rely on.

This makes more sense. I still wonder if the author isn't just effectively recreating Apache Iceberg manually here.

I intentionally kept it lightweight. Just Parquet files + simple partitioning + commits on Hugging Face. That already covers most of what I need, without introducing a heavier stack or extra dependencies.

Also, I wanted something that is easy to consume anywhere. With this setup, you can point DuckDB or Polars directly at the data and start querying, no catalog or special tooling required.

Are they paying for the repo space, I wonder?

someones paying to keep name dropping Iceberg(tm)

Weird accusation. Iceberg is an Apache project. I don’t think anyone gets paid when you use it so not sure what the benefit of shilling would be. It is just a table format that’s well suited for this purpose. I would expect any professional to make a similar recommendation.

So they are sharding by time/day?

I have a similar project right now where I am scraping a dataset that is only ever offering the current state. I am trying to preserve the history of this dataset and was thinking of using the same strategy. If anyone has experience or pointers in how to best add time as a dimension to an existing generic dataset, I'd love to read about it.

"The dataset is organized as one Parquet file per calendar month, plus 5-minute live files for today's activity. Every 5 minutes, new items are fetched from the source and committed directly as a single Parquet block. At midnight UTC, the entire current month is refetched from the source as a single authoritative Parquet file, and today's individual 5-minute blocks are removed from the today/ directory."

So it's not really one big file getting replaced all the time. Though a less extreme variation of that is happening day to day.

Parquet is a very efficient storage approach. Data interfaces tend to treat paths as partitions, if logical.

Was thinking the same thing. probably once a day would be more than enough. if you really want a minute by minute probably a delta file from the previous day should be more than enough.

Reply on news.ycombinator.com