An embedded storage engine for append-only data

In your process like DuckDB, and what it writes is an Iceberg table.

append() returns once the row is durable, and a query a moment later sees it.

litelink is an open-source embedded storage engine: what DuckDB is to query execution, litelink is to the durable write path. It runs inside your process, with no server, daemon or catalog service, and the files it writes are the product. They're Iceberg v2 tables on local disk and in object storage: the Parquet a row is sealed into is the Parquet DuckDB, or any other Iceberg engine, reads, with no export step in between.

SQLite buffer          durable on commit. unsealed rows only.
      │  seal at target_seal_size
      ▼
local Iceberg table    a rolling window. reads land here.
      │  sync: upload data files, register into the archive
      ▼
remote Iceberg table   full history, on S3.

It's built for the thing every capture pipeline hand-rolls badly: getting a stream of observations onto disk durably, into well-sized Parquet, and eventually into object storage. Doing that by hand goes wrong the same way every time: one production capture system had 125,884 objects, 62.5% of them under 16 KiB, Parquet files at 2 rows each, a compaction routine nothing ever scheduled, and an in-memory buffer a SIGKILL emptied.

Status: early. All three tiers work, and a log survives losing its machine. Read what it is not and not implemented yet first.

The Parquet is the product

The usual shape is a write path in one system and an analytical store in another, with a job copying between them. Here they are one store with tiers: rows land in the SQLite buffer, seal into Parquet behind it, and reads span both, so no read on the hot path touches the network. Every other machine reads the archive, with litelink or with nothing from it:

import duckdb
import litelink

# Written, durable on return.
log = litelink.open("data", "trades")
log.append({"trade_id": 624438572, "event_ts": 1787772776240000,
            "price": 78501.62, "amount": 0.0076})

# Read on the same box, across the buffer and the local table.
log.sql("SELECT count(*), max(price) FROM log").read_all()

# Read from another box, over the archive — no local root, no catalog service.
with litelink.snapshot("trades", archive="s3://bucket/prefix") as reader:
    reader.scan(where="price > 78000").read_all()

# Or with any Iceberg engine, and litelink not installed at all.
duckdb.sql("""
    SELECT count(*), max(price)
    FROM iceberg_scan('s3://bucket/prefix/trades',
                      version_name_format = '%s%s.metadata.json')
""")

How it works

• Iceberg is used, not reimplemented. Manifests, per-file column statistics, schema with field IDs, and atomic snapshot commits all come from it.
• The library owns exactly one column, litelink_offset — monotonic, never reused. It is the boundary mechanism between tiers. Everything else is the caller's schema.
• Parts are sealed once and never rewritten. Rewriting a growing partition costs ~144x write amplification and buys nothing, because the local WAL already made the row durable.
• Read boundaries come from committed table state, never from a stored flag — so no seal window can double-count or drop.
• Sizing is two targets, not one. A seal wants to be small, because the buffer is what a hot read scans; a file wants to be large, because per-file overhead dominates scans and uploads. Compaction bridges them, on local disk, at 8× the seal size by default.

Read performance is the cost of reading Parquet, plus ~4 ms of fixed overhead. The reasoning and the measurements are in docs/SPEC.md; just bench reruns them on your hardware.

Install

pip install litelink        # or: uv add litelink

Nothing else is required — no producer, no credentials, no maintainer process, no container. Object storage, WAL replication and cross-machine reads are all opt-in, and each is one call.

Wheels for Linux and macOS on x86-64 and arm64 carry a checksum-verified litestream and the DuckDB extensions litelink loads, so a box with no egress still reads, writes and restores. That costs ~124 MB. Run python -m litelink to check a machine before you rely on it; see docs/RUNTIME.md for anywhere else.

API

litelink.new(root, name, *, schema, sort_by=None, config=None, archive=None,
             s3=None, include_archive=False, start_offset=1)     -> WriteHandle
litelink.open(root, name, *, s3=None, include_archive=False)       -> WriteHandle
litelink.open(root, name, *, read_only=True, ...)                  -> LocalReadHandle
litelink.snapshot(name, *, archive, s3=None, include_wal=False, ...) -> RemoteReadHandle
litelink.restore(root, name, *, archive, s3=None, ...)             -> WriteHandle
litelink.validate_row(schema, row)                                 # raises as append would
litelink.preflight(...)                                            # what python -m litelink runs

# Every handle reads:
    log.scan(*, columns=None, where=None, start_offset=None, end_offset=None)
    log.sql(query)                                  # the log is `log`; both stream Arrow
    log.with_archive() · log.column_statistics(*, tier=None) · log.coverage()
    log.end_offset() · buffered_rows() · table_rows() · table_files() · archived_through()
    log.schema · sort_by · config · archive

# A WriteHandle also writes:
    log.append(row) -> int                          # durable on return
    log.extend(rows) -> list[int]                   # ONE transaction, one fsync
    log.ingest(table_or_reader)                     # Arrow straight to Parquet
    log.seal_due() · log.maintain()                 # seal; compact, evict, expire
    log.sync(*, push_unsettled=False)               # push to the archive
    log.set_config(...) · set_archive(...) · set_sort_by(..., rewrite=True) · add_column(...)

The deliberate choices:

• Handles, not logs. A read handle has no write methods at all, rather than ones that raise, and open(..., read_only=True) is typed so misuse is caught before it runs.
• new takes the shape; open takes none of it. Schema, sort order, config and archive live in the log, so nothing at the call site can disagree with what is on disk.
• The library owns no thread. Nothing seals unless you call seal_due() or maintain(); your loop is the schedule.
• Which tiers a handle reads is fixed when it is built. A scan never starts touching the network because retention happened to run.

Full reference in docs/API.md.

Writing

import litelink
import pyarrow as pa

schema = pa.schema([
    pa.field("trade_id", pa.int64()),
    pa.field("event_ts", pa.int64()),    # microseconds, as the exchange sends them
    pa.field("price", pa.float64()),
    pa.field("amount", pa.float64()),
])

log = litelink.new("data", "trades", schema=schema, sort_by=("event_ts",))

log.append({"trade_id": 624438572, "event_ts": 1787772776240000,
            "price": 78501.62, "amount": 0.0076})     # durable on return
log.extend(group_of_rows)                             # the throughput lever
log.maintain()                                        # compact, evict, expire

extend() commits the whole group in one transaction, so it is one fsync for the batch rather than one per row, and that call size is the write-throughput lever. Loading history is ingest(), which writes Arrow straight to Parquet.

Reading

sql exposes the log as log; scan(where=…, columns=…) is the typed equivalent, and both return a pa.RecordBatchReader rather than a table, so materialising is yours to choose. A reader can open the same log alongside a live writer with litelink.open("data", "trades", read_only=True).

A handle reads local files unless you say otherwise. include_archive=True on the open (or log.with_archive(), which derives a read-only view of an open handle without a second connection) is what reaches object storage:

log.scan(...)                       # local files and the buffer
log.with_archive().scan(...)        # the whole history, including the archive

A handle that cannot reach the archive and finds its local table empty refuses rather than returning the buffer alone. column_statistics() gives every column's bounds and counts from the manifests, without opening a data file.

Reading from another machine

litelink.snapshot is the way in. It resolves the archive's current metadata, handles credentials, and hands back a read handle:

import litelink

with litelink.snapshot("trades", archive="s3://bucket/prefix") as reader:
    reader.sql("SELECT count(*), max(litelink_offset) FROM log").read_all()

archive is the prefix the logs sit under and "trades" is the log, so this reads s3://bucket/prefix/trades/. Credentials come from the environment; pass s3=litelink.S3Options(endpoint=…) for somewhere that is not AWS.

By default this reads the archive alone — no replica, no litestream, no subprocess — and assembles in well under a second. The view is as of the archive frontier, which on a quiet stream can lag indefinitely rather than by the sync interval, because sync holds back a trailing run under target_compact_size. coverage() reports what it can actually serve.

Pass include_wal=True when you want the freshest read there is. With the writer running a WAL sidecar, its buffer.db is restored from the replica and merged with the archive, so you see down to the replication lag rather than to the last sync:

with litelink.snapshot("trades", archive="s3://bucket/prefix", include_wal=True) as reader:
    print(reader.coverage())
    # Coverage(archive=(1, 1928), buffered=(1929, 2100), gap=None, wal_replication=True)

It needs wal_replication on and a sidecar that has shipped; without a replica it raises, which is why it is not the default.

That restore is the expensive part: measured against a 276k-row log, 22 s to assemble against 1.4 s per scan — and it scales with the buffer FILE's size rather than its row count, so a writer whose buffer has grown a large free list makes every reader slower (reclaim_buffer() shrinks it). Assemble once and scan many times; re-entering the with block per query pays it every time.

Either way it is a snapshot, not a subscription — refreshing means assembling another one — and it cannot append: a read handle has no write surface at all, rather than one that raises.

Or any Iceberg engine

The archive is an ordinary Iceberg table that publishes version-hint.text at every commit, so an engine pointed at the prefix resolves the current metadata itself. No catalog service, no local root, no litelink install:

import duckdb

con = duckdb.connect()
con.execute("CREATE SECRET (TYPE s3, PROVIDER credential_chain, REGION 'us-east-1');")

table = con.execute("""
    SELECT count(*), max(litelink_offset)
    FROM iceberg_scan('s3://bucket/prefix/trades',
                      version_name_format = '%s%s.metadata.json')
""").arrow().read_all()

Point it at the table DIRECTORY — <archive>/<name> — not at a metadata JSON. version_name_format is not optional: DuckDB defaults to the Hadoop v%s%s.metadata.json while pyiceberg names its metadata 00003-<uuid>.metadata.json, so the format has to stop prepending the v. credential_chain is the ordinary AWS resolution — profile, instance metadata, SSO; against another endpoint pass KEY_ID, SECRET, ENDPOINT and URL_STYLE 'path' instead. No INSTALL/LOAD is needed — DuckDB autoloads iceberg, avro and httpfs when a query names them, and just bootstrap provisions them ahead of time so the first read is not a download.

Which to reach for. A one-shot query in a script is cheaper this way: snapshot assembles a DuckDB connection, a scratch buffer and an adopted catalog before it can answer anything, and you exit before reusing any of it. Hold a snapshot open and the order reverses — measured on a 200k-row archive, a bounded scan is 0.02 s against 0.40 s for a fresh DuckDB connection, because the connection and the loaded extensions are already there. Assemble once and scan many times, or use the query above.

litelink_offset is monotonic and never reused, so a reader keeps the highest it has seen and asks for what came after — which is how you poll the archive as it grows.

Versioned data

A log can also hold versioned data the way many databases' storage does: append every version of a record, and a delete as a tombstone, and litelink_offset (monotonic and never reused) orders them, so the current state is one query:

SELECT * FROM log
QUALIFY row_number() OVER (PARTITION BY account ORDER BY "litelink_offset" DESC) = 1

What it doesn't do is the rest of MVCC. Nothing merges versions on read or drops superseded ones at compaction, and there is no index, so that query is a scan.

Demos and recovery

just demo-websocket    # a live public feed, one process, ~30 seconds
just demo-capture      # a synthetic feed, driven as hard as you like
just demo-maintain     # in another terminal: seal, compact, evict, expire
just rustfs            # object storage in a container, to add the archive tier
just demo-replicate    # ship the SQLite WAL, to survive losing the machine

Clone the repo for these; just bootstrap sets up the toolchain. Credentials are never written to the log directory — the library reads them from the environment through the ordinary AWS chain, so a profile, instance metadata or SSO all work untouched. litelink.restore(root, name, archive=...) rebuilds a log on another box, reserving an offset window so nothing the dead machine served is reissued.

litelink emits the litestream config; your supervisor runs the binary. Full walkthrough in examples/ and docs/RUNTIME.md.

On disk

One directory per stream, holding everything that stream owns — and the archive prefix mirrors it, so a stream can be copied, replicated or deleted whole in either tier:

data/trades/                     s3://bucket/prefix/trades/
    buffer.db                        _wal/
    catalog.db                           buffer.db/
    archive.db                           catalog.db/
    litestream.yml                       archive.db/
    data/                            data/
        *.parquet                        *.parquet
        compacted/*.parquet              compacted/*.parquet
        ingested/*.parquet               ingested/*.parquet
    metadata/                        metadata/
        *.metadata.json                  *.metadata.json
        *.avro                           *.avro
                                         version-hint.text

Data files sit under the table's own location, so the path an engine reads (s3://bucket/prefix/trades) is the directory that holds both halves of the table.

Upgrading a log written by 0.1.0: see Migrating from 0.1.

What it is not

• Not an OLTP or key-value store. It is append-only, with no update or delete, and a point lookup is ~1,600x slower than an indexed row store, because there is no index to look up: a lookup scans, pruned only by min/max statistics, which are tight on sort_by's leading column and loose elsewhere. Indexes are not implemented yet. It is a local, in-process, real-time analytics store: freshness is sub-second with durability, but "real-time" means fresh, not point-lookup fast.

• Not an unbounded local archive. A seal's cost tracks what the table's metadata holds, so a log that never runs maintain() and never evicts gets slower on the write path over time. maintain() arrests the larger factor; a retention bounds the rest. Numbers and the reasoning are in docs/SPEC.md §13.7.

Not implemented yet

• Indexes for point lookups. A lookup by key scans the tiers, pruned only by min/max statistics, so finding one row costs a scan rather than a seek — least on sort_by's leading column, where the statistics are tight.

• Schema evolution is half built: add_column works, rename_column and drop_column raise NotImplementedError (SPEC §9).

• Blob fields — large payloads that bypass the buffer — are specified and unbuilt; binary columns are carried, for ids and other small values rather than payloads (SPEC §15).

• Payload encoding and local-disk backpressure are open (SPEC §13).

Documentation

• docs/API.md — every public call, on one page
• docs/SPEC.md — the design, and in places still ahead of the code
• docs/RUNTIME.md — writer and maintainer, threads, processes, what crosses between them
• examples/ — the websocket capture, and the synthetic feed with one process per role
• benchmarks/ — the harness, including what litelink costs over raw SQLite
• CONTRIBUTING.md — setup, the gates, and what a good PR here looks like
• SECURITY.md — what to report privately, and what is a known limit instead

Development

just bootstrap          # uv sync + git hooks + DuckDB extensions + litestream
just check              # lint + format-check + typecheck + tests, same as CI
just --list             # the rest

A checkout downloads the DuckDB extensions and litestream that an installed wheel carries, so a contributor provisions what a user does not. Tooling is uv + ruff + ty + pytest; commits follow Conventional Commits, enforced by a hook. See CONTRIBUTING.md.

License

Apache License 2.0 — see LICENSE and NOTICE.