Hyperliquid

Seven Parquet tables from HyperCore: fills, order_statuses, book_diffs, oracle_updates, misc_events, twap_statuses and core_writer_actions, with 156 documented columns. Every row carries Block_Number and Block_Time, so the tables join on block without a separate blocks table. HyperEVM transactions and contract events are not part of this bundle.

Large, and driven by book activity. In a measured 1,000-block slice (about 73 seconds) we counted 142,955 order statuses, 89,241 book diffs and 1,552 fills. Scaled to a day that is roughly 170 million order statuses, 105 million book diffs and 1.8 million fills; exact row counts and sizes come with the delivery manifest, and these figures are estimates.

Hyperliquid's archive holds L2 book snapshots and asset contexts, uploaded roughly monthly, and states that no other historical data is provided that way. This dataset is L4: each book_diffs row names the order and the address behind it, and fills, rejects, cancels, funding and ledger events are all included.

Filter fills on User, then sum ClosedPnl and Fee. Each match appears twice, once per counterparty, so filtering by one User never double counts. Crossed tells you whether that side was the taker, and StartPosition is the signed position before the fill (negative is short).

Plain perps such as HYPE, HIP-3 perps with a dex prefix such as xyz:DRAM, spot pairs by index such as @107, and outcome tokens such as #1890, all in the same stream. Filter on the prefix to separate them.

Because the node emits them that way and we keep the values as emitted. Block_Time, OrderStatus.Time and misc_events.Time are epoch nanoseconds; Fill.Time and Order.Timestamp are epoch milliseconds. Divide by 1e9 or 1e3 respectively when converting to timestamps.

The node emitted a field that the schema does not model yet, and we kept it rather than dropping it. Extra is empty on nearly every row. When a new field appears it lands there first, so nothing is silently lost between schema versions.

With DuckDB: SELECT * FROM read_parquet('fills/*.parquet'). With pandas: pd.read_parquet('fills/'). Parse Px, Sz, Fee and the other money columns with Decimal rather than float; they are stored as decimal strings on purpose.

Yes. These field definitions are the ones Bitquery's Hyperliquid Kafka topics use, so a backtest reading this archive and a production consumer reading the live stream decode with the same code.

Yes. The same tables are sold as three smaller datasets: Hyperliquid Trades (fills and TWAP statuses), Hyperliquid Order Book (order statuses, L4 book diffs and oracle updates) and Hyperliquid Ledger and Funding (ledger events and CoreWriter actions). This bundle is all seven tables in one purchase.

30 days. HyperCore history is kept as a rolling 30-day window, so a purchase covers the 30 days up to yesterday and there is no longer window to buy. For a continuous feed beyond that, ask about the live stream.

As flat Parquet files under a stable S3 layout, one prefix per table and one file per 200-block range, with a JSON manifest listing every file and its sha256. Signed HTTPS links are emailed once the files are prepared. Refreshed daily with T+1 latency; a purchase made today includes everything up to yesterday.

Yes. Every table links to a public sample Parquet file from a real 1,000-block slice, and the sample repository on GitHub shows the first records of each table as readable JSON. No email address required.