FAQ

Is kv0 a database server?

No. It is a library you link into a Rust program or import in Python. Every process that opens a database reads and writes Lance datasets directly.

Why Lance?

It gives atomic, versioned commits on object stores, concurrent appenders that do not conflict, time travel and compaction. kv0 has no storage protocol of its own, so the same code runs on a laptop and on a bucket. The design has the costs.

How durable is a write?

It survives a killed process once put returns (it is in the local write-ahead log). With the default wal_sync="batch", a power loss can lose about the last 100 ms; "always" fsyncs every write. Losing the machine loses writes not yet committed; flush() or sync=True commits before returning. See Concepts.

Can several processes write at once?

Yes, on one machine or many. Appends do not conflict; when two writers write one key, the higher sequence number wins everywhere. There are no transactions across handles.

How much memory does it use?

The keys of the live data, about 100 bytes per entry plus key and small value; values over 4 KiB stay in Lance. 10M keys use about 1 GB. The benchmarks have the numbers.

How long does opening take?

With a snapshot from a previous close on the same machine, about 70 ms per million keys. Without one, a full scan: about 0.5 s per million keys.

Does it support transactions, merge operators or column families?

WriteBatch is atomic. There are no multi-key transactions with reads. merge is accepted and behaves as a put (last write wins); column-family arguments are accepted and every key lives in one keyspace.

What does it cost on S3?

A commit is a few requests and takes tens to hundreds of milliseconds, so kv0 buffers writes and commits in batches. Reads are served from memory; a refresh is one listing plus the new data.

Where are the benchmarks from?

One machine, five repetitions per cell, a stated significance rule, and every regression reported in the same tables. Benchmarks.