Kgoss/hacker-news

Hacker News - Complete Archive

Every Hacker News item since 2006, live-updated every 5 minutes

What is it?

This dataset contains the complete Hacker News archive: every story, comment, Ask HN, Show HN, job posting, and poll ever submitted to the site. Hacker News is one of the longest-running and most influential technology communities on the internet, operated by Y Combinator since 2007. It has become the de facto gathering place for founders, engineers, researchers, and technologists to share and discuss what matters in technology.

The archive currently spans from 2006-10 to 2026-04-18 20:15 UTC, with 47,747,219 items committed. New items are fetched every 5 minutes and committed directly as individual Parquet files through an automated live pipeline, so the dataset stays current with the site itself.

We believe this is one of the most complete and regularly updated mirrors of Hacker News data available on Hugging Face. The data is stored as monthly Parquet files sorted by item ID, making it straightforward to query with DuckDB, load with the datasets library, or process with any tool that reads Parquet.

What is being released?

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.

data/
  2006/2006-10.parquet       first month with HN data
  2006/2006-12.parquet
  2007/2007-01.parquet
  ...
  2026/2026-04.parquet   most recent complete month
  2026/2026-04.parquet  current month, ongoing til 2026-04-17
today/
  2026/04/18/00/00.parquet  5-min live blocks (YYYY/MM/DD/HH/MM.parquet)
  2026/04/18/00/05.parquet
  ...
  2026/04/18/20/15.parquet  most recent committed block
stats.csv                    one row per committed month
stats_today.csv              one row per committed 5-min block

Along with the Parquet files, we include stats.csv which tracks every committed month with its item count, ID range, file size, fetch duration, and commit timestamp. This makes it easy to verify completeness and track the pipeline's progress.

Breakdown by today

The chart below shows items committed to this dataset by hour today (2026-04-18, 5,976 items across 21 hours, last updated 2026-04-18 20:20 UTC).

  00:00  ██████████████████░░░░░░░░░░░░  298
  01:00  ██████████████░░░░░░░░░░░░░░░░  234
  02:00  ████████████░░░░░░░░░░░░░░░░░░  200
  03:00  ████████████░░░░░░░░░░░░░░░░░░  192
  04:00  ████████████░░░░░░░░░░░░░░░░░░  193
  05:00  ████████████░░░░░░░░░░░░░░░░░░  197
  06:00  ████████████░░░░░░░░░░░░░░░░░░  202
  07:00  ████████████████░░░░░░░░░░░░░░  260
  08:00  █████████████░░░░░░░░░░░░░░░░░  220
  09:00  ████████████░░░░░░░░░░░░░░░░░░  196
  10:00  ███████████████░░░░░░░░░░░░░░░  239
  11:00  ███████████████░░░░░░░░░░░░░░░  245
  12:00  ██████████████████░░░░░░░░░░░░  288
  13:00  █████████████████████░░░░░░░░░  342
  14:00  ██████████████████████░░░░░░░░  357
  15:00  █████████████████████████░░░░░  411
  16:00  ████████████████████████████░░  452
  17:00  ████████████████████████████░░  453
  18:00  ████████████████████████░░░░░░  389
  19:00  ██████████████████████████████  477
  20:00  ████████░░░░░░░░░░░░░░░░░░░░░░  131

Breakdown by year

The chart below shows items committed to this dataset by year.

  2006  █░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  62
  2007  █░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  93.8K
  2008  ██░░░░░░░░░░░░░░░░░░░░░░░░░░░░  320.9K
  2009  ███░░░░░░░░░░░░░░░░░░░░░░░░░░░  608.4K
  2010  ██████░░░░░░░░░░░░░░░░░░░░░░░░  1.0M
  2011  ████████░░░░░░░░░░░░░░░░░░░░░░  1.4M
  2012  ██████████░░░░░░░░░░░░░░░░░░░░  1.6M
  2013  █████████████░░░░░░░░░░░░░░░░░  2.0M
  2014  ███████████░░░░░░░░░░░░░░░░░░░  1.8M
  2015  █████████████░░░░░░░░░░░░░░░░░  2.0M
  2016  ████████████████░░░░░░░░░░░░░░  2.5M
  2017  █████████████████░░░░░░░░░░░░░  2.7M
  2018  ██████████████████░░░░░░░░░░░░  2.8M
  2019  ████████████████████░░░░░░░░░░  3.1M
  2020  ████████████████████████░░░░░░  3.7M
  2021  ███████████████████████████░░░  4.2M
  2022  █████████████████████████████░  4.4M
  2023  ██████████████████████████████  4.6M
  2024  ████████████████████████░░░░░░  3.7M
  2025  █████████████████████████░░░░░  3.9M
  2026  ████████░░░░░░░░░░░░░░░░░░░░░░  1.4M

How to download and use this dataset

You can load the full dataset, a specific year, or even a single month. The dataset uses the standard Hugging Face Parquet layout, so it works out of the box with DuckDB, the datasets library, pandas, and huggingface_hub.

Using DuckDB

DuckDB can read Parquet files directly from Hugging Face without downloading anything first. This is the fastest way to explore the data:

The type column is stored as a small integer: 1 = story, 2 = comment, 3 = poll, 4 = pollopt, 5 = job. The "by" column (author username) must be quoted in DuckDB because by is a reserved keyword.

-- Top 20 highest-scored stories of all time
SELECT id, title, "by", score, url, time
FROM read_parquet('hf://datasets/open-index/hacker-news/data/*/*.parquet')
WHERE type = 1 AND title != ''
ORDER BY score DESC
LIMIT 20;

-- Monthly submission volume for a specific year
SELECT
    strftime(time, '%Y-%m') AS month,
    count(*) AS items,
    count(*) FILTER (WHERE type = 1) AS stories,
    count(*) FILTER (WHERE type = 2) AS comments
FROM read_parquet('hf://datasets/open-index/hacker-news/data/2024/*.parquet')
GROUP BY month
ORDER BY month;

-- Most discussed stories by total comment count
SELECT id, title, "by", score, descendants AS comments, url
FROM read_parquet('hf://datasets/open-index/hacker-news/data/2025/*.parquet')
WHERE type = 1 AND descendants > 0
ORDER BY descendants DESC
LIMIT 20;

-- Who posts the most Ask HN questions?
SELECT "by", count(*) AS posts
FROM read_parquet('hf://datasets/open-index/hacker-news/data/*/*.parquet')
WHERE type = 1 AND title LIKE 'Ask HN:%'
GROUP BY "by"
ORDER BY posts DESC
LIMIT 20;

-- Track how often a topic appears on HN over time
SELECT
    extract(year FROM time) AS year,
    count(*) AS mentions
FROM read_parquet('hf://datasets/open-index/hacker-news/data/*/*.parquet')
WHERE type = 1 AND lower(title) LIKE '%rust%'
GROUP BY year
ORDER BY year;

-- Top linked domains, year over year
SELECT
    extract(year FROM time) AS year,
    regexp_extract(url, 'https?://([^/]+)', 1) AS domain,
    count(*) AS stories
FROM read_parquet('hf://datasets/open-index/hacker-news/data/*/*.parquet')
WHERE type = 1 AND url != ''
GROUP BY year, domain
QUALIFY row_number() OVER (PARTITION BY year ORDER BY stories DESC) <= 5
ORDER BY year, stories DESC;

Using datasets

from datasets import load_dataset

# Stream the full history without downloading everything first
ds = load_dataset("open-index/hacker-news", split="train", streaming=True)
for item in ds:
    print(item["id"], item["type"], item["title"])

# Load a specific year into memory
ds = load_dataset(
    "open-index/hacker-news",
    data_files="data/2024/*.parquet",
    split="train",
)
print(f"{len(ds):,} items in 2024")

# Load today's live blocks (updated every 5 minutes)
ds = load_dataset(
    "open-index/hacker-news",
    name="today",
    split="train",
    streaming=True,
)

Using huggingface_hub

from huggingface_hub import snapshot_download

# Download only 2024 data (about 1.5 GB)
snapshot_download(
    "open-index/hacker-news",
    repo_type="dataset",
    local_dir="./hn/",
    allow_patterns="data/2024/*",
)

For faster downloads, install pip install huggingface_hub[hf_transfer] and set HF_HUB_ENABLE_HF_TRANSFER=1.

Using the CLI

# Download a single month
huggingface-cli download open-index/hacker-news \
    data/2024/2024-01.parquet \
    --repo-type dataset --local-dir ./hn/

Using pandas + DuckDB

import duckdb

conn = duckdb.connect()

# Score distribution: what does a "typical" HN story look like?
# type=1 is story (stored as integer: 1=story, 2=comment, 3=poll, 4=pollopt, 5=job)
df = conn.sql("""
    SELECT
        percentile_disc(0.50) WITHIN GROUP (ORDER BY score) AS p50,
        percentile_disc(0.90) WITHIN GROUP (ORDER BY score) AS p90,
        percentile_disc(0.99) WITHIN GROUP (ORDER BY score) AS p99,
        percentile_disc(0.999) WITHIN GROUP (ORDER BY score) AS p999
    FROM read_parquet('hf://datasets/open-index/hacker-news/data/*/*.parquet')
    WHERE type = 1
""").df()
print(df)

Dataset statistics

You can query the per-month statistics directly from the stats.csv file included in the dataset:

SELECT * FROM read_csv_auto('hf://datasets/open-index/hacker-news/stats.csv')
ORDER BY year, month;

The stats.csv file tracks each committed month with the following columns:

Column	Description
year, month	Calendar month
lowest_id, highest_id	Item ID range covered by this file
count	Number of items in the file
dur_fetch_s	Seconds to fetch from the data source
dur_commit_s	Seconds to commit to Hugging Face
size_bytes	Parquet file size on disk
committed_at	ISO 8601 timestamp of when this month was committed

Content breakdown

Hacker News has five item types. The vast majority of content is comments, followed by stories (which include Ask HN, Show HN, and regular link submissions). Jobs, polls, and poll options make up a small fraction.

Type	Count	Share
comment	41,616,741	87.2%
story	6,079,812	12.7%
job	18,094	0.0%
poll	2,242	0.0%
pollopt	15,463	0.0%

Of all stories submitted to Hacker News, 84.6% link to an external URL. The rest are text-only posts: Ask HN questions, Show HN launches, and other self-posts where the discussion itself is the content.

The average story generates 23.8 comments in its discussion thread. The most-discussed story of all time received 9,275 comments, which gives a sense of how deep conversations can go on particularly controversial or interesting topics.

Story scores

Scores on Hacker News follow a steep power law. Most stories receive only a few points, but a small number break out and reach the front page with hundreds or thousands of upvotes.

Metric	Value
Average score	1.5
Median score	0
Highest score ever	6,015
Stories with 100+ points	175,933
Stories with 1,000+ points	2,169

The median score of 0 reflects the fact that many stories are submitted but never gain traction. However, the long tail is where things get interesting: over 6,079,812 stories have been submitted, and the top 0.03% (those with 1,000+ points) represent the content that defined conversations across the technology industry.

Most-shared domains

The domains most frequently linked from Hacker News stories tell a clear story about what the community values. GitHub dominates, reflecting HN's deep roots in open source and software development. Major publications like the New York Times and Ars Technica show the community's interest in journalism and long-form analysis.

#	Domain	Stories
1	github.com	200,995
2	www.youtube.com	135,566
3	medium.com	124,707
4	www.nytimes.com	77,999
5	en.wikipedia.org	54,655
6	techcrunch.com	54,281
7	twitter.com	50,955
8	arstechnica.com	47,259
9	www.theguardian.com	44,557
10	www.bloomberg.com	37,959

Most active story submitters

These are the users who have submitted the most stories over the lifetime of Hacker News. Many of them have been active for over a decade, consistently curating and sharing content with the community.

#	User	Stories
1	rbanffy	36,901
2	Tomte	26,259
3	tosh	24,273
4	bookofjoe	20,785
5	mooreds	20,641
6	pseudolus	19,974
7	PaulHoule	19,207
8	todsacerdoti	18,887
9	ingve	17,117
10	thunderbong	16,130
11	jonbaer	14,195
12	rntn	13,410
13	doener	12,971
14	Brajeshwar	12,738
15	LinuxBender	11,058

How it works

The pipeline is built in Go and uses DuckDB for Parquet conversion. Historical data is sourced from ClickHouse; live data is fetched directly from the HN Firebase API.

Historical backfill. The pipeline iterates through every month from October 2006 to the most recent complete month. For each month, it queries the ClickHouse source with a time-bounded SQL query, exports the result as a Parquet file sorted by id using DuckDB with Zstandard compression at level 22, and commits it to this repository along with an updated stats.csv and README.md. Months already tracked in stats.csv are skipped, making the process fully resumable.

Live polling. Every 5 minutes, the pipeline calls the HN Firebase API to fetch new items by ID range. Items are grouped into their 5-minute time windows, written as individual Parquet files at today/YYYY/MM/DD/HH/MM.parquet using DuckDB, and committed to Hugging Face immediately. Using the HN API directly means live blocks reflect real-time data with no indexing lag.

Day rollover. At midnight UTC, the entire current month is refetched from the ClickHouse source in a single query and written as an authoritative Parquet file. Today's individual 5-minute blocks are deleted from the repository in the same atomic commit. Refetching instead of merging ensures the monthly file is always complete and deduplicated, regardless of any local state.

Thanks

The data in this dataset comes from the ClickHouse Playground, a free public SQL endpoint maintained by ClickHouse, Inc. that mirrors the official Hacker News Firebase API. ClickHouse uses Hacker News as one of their canonical demo datasets. Without their public endpoint, building and maintaining a complete, regularly updated archive like this would not be practical.

The original content is created by the Hacker News community and is operated by Y Combinator. This is an independent mirror and is not affiliated with or endorsed by Y Combinator or ClickHouse, Inc.

Dataset card for Hacker News - Complete Archive

Dataset summary

This dataset is a complete mirror of the Hacker News archive, sourced from the ClickHouse Playground which itself mirrors the official HN Firebase API. The data covers every item ever posted to the site, from the earliest submissions in October 2006 through today.

The dataset is intended for research, analysis, and training. Common use cases include:

• Language model pretraining and fine-tuning on high-quality technical discussions
• Sentiment and trend analysis across two decades of technology discourse
• Community dynamics research on one of the internet's most influential forums
• Information retrieval benchmarks using real-world questions and answers
• Content recommendation and ranking model development

Dataset structure

Data instances

Here is an example item from the dataset. This is a story submission with a link to an external URL:

{
  "id": 1,
  "deleted": 0,
  "type": 1,
  "by": "pg",
  "time": "2006-10-09T18:21:51+00:00",
  "text": "",
  "dead": 0,
  "parent": 0,
  "poll": 0,
  "kids": [15, 234509, 487171],
  "url": "http://ycombinator.com",
  "score": 57,
  "title": "Y Combinator",
  "parts": [],
  "descendants": 0,
  "words": ["y", "combinator"]
}

And here is a comment, showing how discussion threads are connected via the parent field:

{
  "id": 15,
  "deleted": 0,
  "type": 2,
  "by": "sama",
  "time": "2006-10-09T19:51:01+00:00",
  "text": "\"the way to get good software is to find ...",
  "dead": 0,
  "parent": 1,
  "poll": 0,
  "kids": [17],
  "url": "",
  "score": 0,
  "title": "",
  "parts": [],
  "descendants": 0,
  "words": []
}

Data fields

Every Parquet file shares the same schema, matching the HN API item format:

Column	Type	Description
id	uint32	Unique item ID, monotonically increasing across the entire site
deleted	uint8	1 if the item was soft-deleted by its author or by moderators, 0 otherwise
type	int8	Item type as an integer: 1=story, 2=comment, 3=poll, 4=pollopt, 5=job
by	string	Username of the author who created this item. Note: by is a reserved word in DuckDB and must be quoted as "by"
time	timestamp	When the item was created, in UTC
text	string	HTML body text. Used for comments, Ask HN posts, job listings, and polls
dead	uint8	1 if the item was flagged or killed by moderators, 0 otherwise
parent	uint32	The ID of the parent item. For comments, this points to either a story or another comment
poll	uint32	For poll options (pollopt), the ID of the associated poll
kids	list<uint32>	Ordered list of direct child item IDs (typically comments)
url	string	The external URL for link stories. Empty for text posts and comments
score	int32	The item's score (upvotes minus downvotes)
title	string	Title text for stories, jobs, and polls. Empty for comments
parts	list<uint32>	For polls, the list of associated poll option item IDs
descendants	int32	Total number of comments in the entire discussion tree below this item
words	list<string>	Tokenized words extracted from the title and text fields

Data splits

The default configuration includes all historical monthly Parquet files. If you only need today's latest items, use the today configuration which includes only the 5-minute live blocks for the current day.

You can also load individual years or months by specifying data_files:

# Load just January 2024
ds = load_dataset("open-index/hacker-news", data_files="data/2024/2024-01.parquet", split="train")

# Load all of 2024
ds = load_dataset("open-index/hacker-news", data_files="data/2024/*.parquet", split="train")

Dataset creation

Curation rationale

Hacker News is one of the richest sources of technical discussion on the internet, but accessing the full archive programmatically has historically required either scraping the Firebase API item-by-item or working with incomplete third-party dumps. This dataset provides the complete archive in a standard, efficient format that anyone can query without setting up infrastructure.

By publishing on Hugging Face with Parquet files, the data becomes immediately queryable with DuckDB (via hf:// paths), streamable with the datasets library, and downloadable in bulk. The 5-minute live update pipeline means researchers always have access to near-real-time data.

Source data

All data is sourced from the ClickHouse Playground, a public SQL endpoint maintained by ClickHouse that mirrors the official Hacker News Firebase API. The ClickHouse mirror is widely used for analytics demonstrations and contains the complete dataset.

The pipeline queries the ClickHouse endpoint month-by-month, exports each month as a Parquet file using DuckDB with Zstandard compression at level 22, and commits it to this Hugging Face repository. Already-committed months are tracked in stats.csv and skipped on subsequent runs, making the process fully resumable.

Data processing steps

The pipeline runs in three modes:

1. Historical backfill. Iterates through every month from October 2006 to the most recent complete month. For each month, it runs a SQL query against the ClickHouse source, writes the result as a Parquet file sorted by id, and commits it to Hugging Face along with an updated stats.csv and README.md.

2. Live polling. After the historical backfill completes, the pipeline polls the HN Firebase API every 5 minutes for new items. It fetches all items with IDs greater than the last committed watermark, groups them into 5-minute time windows by item timestamp, and writes each window as a today/YYYY/MM/DD/HH/MM.parquet file committed to Hugging Face immediately. The HN API provides real-time data with no indexing lag.

3. Day rollover. At midnight UTC, the entire current month is refetched from the ClickHouse source in a single query and written as a fresh, authoritative Parquet file. Today's individual 5-minute blocks are deleted from the repository in the same atomic commit. This approach is more reliable than merging local blocks — the result is always complete and deduplicated, sourced directly from the origin.

All Parquet files use Zstandard compression at level 22 and are sorted by id for efficient range scans. No filtering, deduplication, or transformation is applied to the data beyond what the source provides.

Personal and sensitive information

This dataset contains usernames (by field) and user-generated text content (text, title fields) as they appear on the public Hacker News website. No additional PII processing has been applied. The data reflects what is publicly visible on news.ycombinator.com.

If you find content in this dataset that you believe should be removed, please open a discussion on the Community tab.

Considerations for using the data

Social impact

By providing the complete Hacker News archive in an accessible format, we hope to enable research into online community dynamics, technology trends, and the evolution of technical discourse. The dataset can serve as training data for language models that need to understand technical discussions, or as a benchmark for information retrieval and recommendation systems.

Discussion of biases

Hacker News has a well-documented set of community biases. The user base skews heavily toward software engineers, startup founders, and technology enthusiasts based in the United States. Topics related to Silicon Valley, programming languages, startups, and certain political viewpoints tend to receive disproportionate attention and engagement.

The moderation system (flagging, vouching, and moderator intervention) shapes what content survives and what gets killed. Stories and comments that violate community norms are flagged as dead, but this moderation reflects the values of the existing community rather than any objective standard.

We have not applied any additional filtering or quality scoring to the data. All items, including deleted and dead items, are preserved exactly as they appear in the source.

Known limitations

• type is an integer. The item type is stored as a TINYINT enum: 1=story, 2=comment, 3=poll, 4=pollopt, 5=job. When writing DuckDB queries, use WHERE type = 1 for stories rather than WHERE type = 'story'.
• by is a reserved keyword in DuckDB. Always quote it with double quotes: "by".
• deleted and dead are integers. They are stored as 0/1 rather than booleans.
• Comment text is HTML. The text field contains raw HTML as stored by HN, not plain text. You may need to strip tags depending on your use case.
• Deleted items have sparse fields. When an item is deleted, most fields become empty, but the id and deleted flag are preserved.
• Scores are point-in-time snapshots. The score reflects the value at the time the ClickHouse mirror last synced, not necessarily the final score.
• No user profiles. This dataset contains items only, not user profiles (karma, bio, etc.).
• Code content is HTML-escaped. Code snippets in comments use HTML entities and <code> tags rather than Markdown formatting.

Additional information

Licensing

The dataset is released under the Open Data Commons Attribution License (ODC-By) v1.0. The original content is subject to the rights of its respective authors. Hacker News data is provided by Y Combinator.

This is an independent community mirror. It is not affiliated with or endorsed by Y Combinator.

Contact

For questions, feedback, or issues, please open a discussion on the Community tab.

Last updated: 2026-04-18 20:20 UTC