PowDB

PowDB is a pure-Rust embedded database whose query language returns shaped results: one row per parent with its children nested inside, no join fan-out and no JSON text round-trip. Its compiled execution engine measures 3-7x SQLite on aggregates and 1-3.7x on filtered scans, and roughly 16x slower than SQLite on indexed point lookups.

Performance -- compiled byte-level predicates, zero-copy mmap scans, and a plan cache with literal substitution. Filter and aggregate paths skip full row decoding.
Platform -- pure-Rust engine (powdb, powdb-storage, powdb-query pull no C at all), embeddable and server modes, installed with a single cargo install on Linux and macOS. A built binary needs nothing installed beside it, but building powdb-server or powdb-cli from source does need a C toolchain and cmake for their TLS stack; see Install. Windows is not supported (the storage engine's mmap scan path is Unix-only); see Platform support.
DX -- PowQL is the front door: a left-to-right pipeline syntax that reads like an iterator chain.

Website: zvn-dev.github.io/powdb

Evaluating PowDB? Start with the honest comparison: PowDB vs SQLite -- when to use which. What an upgrade may and may not break: docs/STABILITY.md.

What PowDB is for

PowDB is a single-writer embedded engine with truly parallel reads. Every writer takes the whole write-admission gate; there is no MVCC. That shape (a permanent, deliberate one) decides where the engine shines and where it does not.

Reach for PowDB when:

Single-writer embedded app state in Rust or Node. One process owns the data, reads run in parallel, and you want the engine in-process (Rust crate or the @zvndev/powdb-embedded Node addon) with lossless typed results and injection-inert parameters.
Local agent or tool memory. Local-first and single-process. Freshness comes from swapping in a new snapshot in seconds, not from tailing a live feed.
Read-only edge snapshot serving. Restore a backup and serve the directory read-only from N processes, with no write gate at all; refresh by restoring a newer snapshot beside it and swapping.
Per-tenant, process-isolated databases. One directory and one writer per tenant, rather than many tenants contending on one shared database.
CI and test databases. Fast to create, cheap to throw away.
Bulk ingest plus read-heavy internal tools. Batch-load in a transaction, then serve dashboards and internal queries off a read-mostly database.

Use something else when:

Many concurrent clients share one read-write database. That is Postgres's home turf. PowDB serializes writers through one admission gate, so on a shared read-write database at concurrency, read latency amplifies through that gate. Use Postgres.
You need live replication or sync across nodes. Use Turso.
Your workload is analytical column-crunching over one big dataset. Use DuckDB.

For the concurrency numbers behind the boundary above (single-request cost versus behavior at concurrency 10), decomposed with a reproducible script, see docs/benchmarks/concurrency-decomposition.md.

How it works

Compiled predicate engine. Filter expressions on integer columns are compiled into branch-free, byte-level operations that run directly against the encoded row bytes. The executor pattern-matches on Filter(SeqScan) plan shapes and dispatches to fast paths that never decode columns they don't need. On aggregate workloads this is where the 3-7x SQLite wins come from; on scan-shaped workloads the same machinery buys 1-3.7x. It does nothing for point lookups, where PowDB is roughly 16x slower than SQLite.

Plan cache + tight planner-executor contract. The planner is pure (no catalog access) and produces a canonical PlanNode tree; the cache hashes the canonical shape with FNV-1a and substitutes literals at lookup time, so repeat queries skip lex/parse/plan entirely. Range scans without a matching index are lowered to Filter(SeqScan) at execution time, keeping the planner stateless and the executor's fast paths fireable.

Zero-copy mmap scans. Heap files are memory-mapped and scanned via try_for_each_row_raw, a zero-syscall, zero-allocation iterator over raw row bytes. Combined with the compiled predicates, a full scan + filter + count never copies a row.

PowQL -- the front door. PowQL replaces SQL's inside-out clause structure with a left-to-right pipeline. You name the table, chain operations, and project fields, all in reading order.

Task	SQL	PowQL
Filter + project	`SELECT name, age FROM User WHERE age > 25`	`User filter .age > 25 { .name, .age }`
Sort + limit	`SELECT * FROM User ORDER BY age DESC LIMIT 10`	`User order .age desc limit 10`
Aggregate with filter	`SELECT AVG(age) FROM User WHERE city = 'NYC'`	`avg(User filter .city = "NYC" { .age })`
Group + having	`SELECT status, COUNT(name) FROM User GROUP BY status HAVING COUNT(name) > 5`	`User group .status having count(.name) > 5 { .status, n: count(.name) }`
Follow a relationship	`SELECT o.total, u.name FROM Order o JOIN User u ON o.user_id = u.id`	`Order as o { o.total, o.user.name }` (after `link Order.user -> User on user_id = id`)

PowQL uses .field dot syntax for column references, := for assignments, and "double quotes" for strings. The pipeline reads like a sentence: "User, filter age greater than 25, order by name, limit 10, give me name and age."

Nested projections (correlated children as a native array, one row per parent) and entity links (declare a relationship once, then traverse it by name) are the two PowQL-only spellings. They are spellings, not capabilities SQL lacks: stock SQLite reproduces the same shaped output, [] for a childless parent included, with a correlated subquery and json_group_array(json_object(...)). What PowQL adds over that is real but narrow: one line instead of three, the correlation declared once in the catalog instead of retyped in every query, and a value that reaches the client as PJ1 binary rather than as JSON text it has to parse back.

Where PowQL changes the answer, not the spelling: aggregates over a join. A one-to-many join repeats the parent once per child, so an average over the parent's column is inflated by the fan-out. Three accounts (10, 10, 40) with 4, 1 and 1 orders average 20 by themselves and 15 through the join. SQL gives you the 15 unless you know to write a DISTINCT subquery; PowQL's aggregates are symmetric by default and give you the 20, with avg(raw a.balance) as the explicit opt-out when you do want the joined-row number. This is the one place PowDB is correct by default where the obvious SQL is quietly wrong, and it is documented in Grouped aggregates over joins. PowDB's own SQL frontend keeps SQL's raw semantics on purpose.

Already think in SQL? Since v0.5.0 PowDB also accepts a supported subset of SQL through a frontend that lowers to the same PowQL plan tree (and shares the plan cache), see docs/SQL.md. PowQL remains the native, fastest path.

Full language reference: docs/POWQL.md | SQL frontend: docs/SQL.md | Getting started: docs/getting-started.md | Backup & restore: docs/backup-and-restore.md | Driver/ORM implementers: docs/integrations/powql-for-drivers.md | Wire error codes: docs/errors.md | CLI reference (--exec, --exec-file, --sql, --format, REPL meta-commands): crates/cli/README.md | Stability and upgrade policy: docs/STABILITY.md

Install

# Embedded in a Rust project (the engine as a library; needs no C toolchain)
cargo add powdb

# CLI + server from crates.io (Rust 1.93+; needs a C toolchain and cmake, see below)
cargo install powdb-cli
cargo install powdb-server

# TypeScript client (Node 18+): version is kept in lockstep with the workspace by scripts/check-version-consistency.sh
npm install @zvndev/powdb-client

# In-process Node addon: embed the engine directly, no server (prebuilt for macOS arm64, Linux x64-gnu, Linux arm64-gnu ONLY; no source fallback. Elsewhere `require()` throws an error coded `unsupported_platform` that names the three supported targets: use @zvndev/powdb-client there)
npm install @zvndev/powdb-embedded

# Prebuilt binaries (linux x86_64, macos aarch64)
# https://github.com/ZVN-DEV/powdb/releases/latest

# Docker
docker pull ghcr.io/zvn-dev/powdb:latest

# Or build from source
git clone https://github.com/ZVN-DEV/powdb
cd powdb
cargo build --release

Requires Rust 1.93+. This builds all crates: the storage engine, query engine, TCP server, CLI, and benchmarks.

Embedded in Rust

The powdb crate is the whole engine as a library, no server and no C toolchain:

use powdb::{Database, QueryResult, Value};

fn main() -> Result<(), powdb::Error> {
    let mut db = Database::open("./data")?;
    db.query("type User { required name: str, age: int }")?;
    db.query(r#"insert User { name := "Ada", age := 36 }"#)?;
    match db.query("count(User)")? {
        QueryResult::Scalar(Value::Int(n)) => assert_eq!(n, 1),
        other => panic!("unexpected: {other:?}"),
    }
    Ok(())
}

Full embedded API docs (read-only opens, memory limits, sync modes, typed results): docs.rs/powdb.

Building powdb-server or powdb-cli requires a C toolchain and cmake, and there is currently no way to opt out. Both reach TLS through tokio-rustls, which pulls aws-lc-sys. Neither crate declares any Cargo features, so --no-default-features is a silent no-op: there is no tls feature to turn off. Making TLS optional is real work we have not done yet.

The whole-workspace cargo build --release above compiles more than the engine: it also builds powdb-compare, our SQLite/Postgres comparison harness, which bundles SQLite from C source via rusqlite and links postgres. Neither ships in any published crate. The engine libraries themselves (powdb, powdb-storage, powdb-query, powdb-auth, powdb-backup, powdb-sync) pull no C at all, so cargo build --release -p powdb needs no C toolchain.

Platform support

Platform	Status
Linux x86_64	Supported (prebuilt binary + `cargo install`)
Linux aarch64	Supported (`cargo install`, multi-arch Docker)
macOS aarch64 (Apple silicon)	Supported (prebuilt binary + `cargo install`)
macOS x86_64 (Intel)	Builds from source; no prebuilt binary
Windows	Not supported. Does not compile.

Windows is not a "build it yourself" case, it is a hard compile failure. The heap's memory-mapped scan path in crates/storage/src/heap.rs calls libc::mmap / libc::munmap and std::os::unix::io::AsRawFd with no platform gate, so cargo check -p powdb-storage --target x86_64-pc-windows-msvc fails with 21 errors before anything else is attempted. Porting it needs a Windows file-mapping backend for that path. The rest of the storage layer is already portable (crates/storage/src/disk.rs handles both platforms), so the gap is narrow, but it is real today.

Apple silicon / arm64

As of v0.10.0, the published ghcr.io/zvn-dev/powdb image is multi-arch (linux/amd64 + linux/arm64), so it runs natively on Apple silicon and ARM servers (e.g. Graviton). Alternatively, install the native binary directly, which builds in under a minute:

cargo install powdb-server

Benchmark: PowDB vs SQLite (100K rows, Apple M5 Max laptop)

PowDB's compiled predicate engine is strongest on read-heavy aggregates, and gives a smaller and more variable win on filtered scans. All 15 workloads are listed, including the four where PowDB ties or loses. For durable write throughput, batch writes in a transaction, see Write throughput & durability.

These are single-request latencies: one query at a time, measuring per-query cost, not throughput under many simultaneous clients. On a shared read-write database at concurrency, reads amplify through the write-admission gate (PowDB has no MVCC); that behavior is decomposed in docs/benchmarks/concurrency-decomposition.md. See also What PowDB is for.

Workload	PowDB	SQLite	Result
Aggregate MIN	221 us	1.70 ms	7.7x faster
Aggregate MAX	217 us	1.47 ms	6.8x faster
Aggregate SUM	234 us	1.45 ms	6.2x faster
Update by primary key	60 ns	272 ns	4.5x faster
Aggregate AVG	455 us	1.70 ms	3.7x faster
Scan + filter + count	380 us	1.40 ms	3.7x faster
Non-indexed point lookup	101 us	319 us	3.2x faster
Scan + filter + sort + limit 10	2.46 ms	6.41 ms	2.6x faster
Multi-column AND filter	1.58 ms	3.21 ms	2.0x faster
Update by filter (10K rows)	2.36 ms	4.54 ms	1.9x faster
Insert single row	380 ns	638 ns	roughly tied
Scan + filter + project top 100	8.1 us	8.9 us	roughly tied
Delete by filter (10K rows)	1.57 ms	1.75 ms	roughly tied
Insert batch (1K rows)	242 ns	214 ns	roughly tied
Indexed point lookup	3.17 us	202 ns	15.7x SLOWER

Reproduce with cargo run --release -p powdb-compare.

The compiled predicate engine avoids full row decoding during scans and aggregates. That is worth 3.7-7.7x on the four aggregates, but only 1.0-3.7x on the four scan-shaped workloads, so read the rows rather than a single headline multiplier. These wins come from compiled predicates and mmap scans, not from PowQL's syntax: the same query written in SQL lowers to the same plan and gets the same numbers.

PowDB loses the indexed point lookup, badly, and by more than we used to publish. Once the index is probed the remaining work is trivial, so nearly the whole 3.17 us is PowDB's own front end (lex, parse, canonicalize, plan-cache lookup) while SQLite amortizes that away with a prepared statement. The previous table put this at 7.9x against an older engine. Nine independent re-measurements of the current engine, across two machines, ranged from 10x to 20x, most of them above 15x. This row got worse and we had been understating it by roughly 2x. If your hot path is "fetch one row by id", SQLite is the better engine and scan throughput will not compensate.

Neither engine fsyncs (PowDB: WalSyncMode::Off, SQLite: :memory:), which isolates query-engine cost from durability cost and is not a durability comparison; for that see Write throughput & durability. Median of 5 runs on an Apple M5 Max (macOS 26.5.1, rustc 1.97.0), commit e3dfa71, 2026-08-15. Re-measured the same way on 2026-09-07 after a large correctness round: every row landed within 11% of the number above and eleven of the fifteen within 3%, so the table is unchanged. The heap allocator rewrite in that round is not visible here because it removes a cost that grows with heap size, and this fixture is too small to pay it; the measurement that does show it is in the changelog. These are laptop numbers, not CI numbers. One caveat specific to the write rows: PowDB writes to a real temp directory while SQLite is :memory:, so insert_single, insert_batch_1k, and delete_by_filter are sensitive to whatever else is touching the disk. Measured under a heavy concurrent build on the same laptop, those rows moved by 30-100x while every other row moved by less than 2x, and two of them changed sign. The table above is from the quietest run we could get, but this machine was not fully idle, so treat the three write rows as the least reliable and re-measure them yourself before relying on them. Full methodology, per-run spread, and what changed in the harness: docs/benchmarks/2026-07-24-wide-bench-snapshot.md.

Write throughput & durability

PowDB is durable by default. The embedded Engine and powdb-server both run in WalSyncMode::Full: an autocommit statement appends to the write-ahead log and fdatasyncs before the call returns, so an acknowledged write has reached stable storage. Reads pay zero fsync cost.

The one thing worth knowing: a single-row insert in autocommit costs one fsync. That caps single-row autocommit at your disk's fsync rate (a few hundred rows/sec on a typical SSD). That is not an engine limit, just the price of durability per statement. The fix is to batch writes in a transaction, which collapses the batch into far fewer of them:

# ~hundreds of rows/sec: one fsync per row
insert User { id := 1, name := "a" }
insert User { id := 2, name := "b" }
...

# ~50x faster, still fully durable: roughly one fsync per 64 rows
begin
insert User { id := 1, name := "a" }
insert User { id := 2, name := "b" }
... thousands of rows ...
commit

Be exact about what a transaction costs, because it is not one fsync. A statement inside begin / commit does not fsync on its own: the durability point is the commit. But the WAL flushes and fsyncs whenever its append buffer reaches 64 records, inside a transaction as much as outside one, and a row is one record. So a 5000-row transaction costs roughly 78 fsyncs, not 5000 and not 1. The same applies to a multi-row insert: it is one statement, but a batch of 5000 rows is still 5000 WAL records.

On a 2026 laptop SSD this is the difference between ~290 rows/sec (autocommit) and ~15,600 rows/sec (one transaction), a 54x speedup, with identical crash-safety either way (the fsync happens per 64 records and at commit, instead of once per row). Always wrap bulk loads and write bursts in a transaction.

The two weaker modes trade that guarantee away, and it is worth being exact about how much:

normal (POWDB_SYNC_MODE=normal, WalSyncMode::Normal) appends to the WAL but does not fsync before acknowledging. A process death loses nothing: the records are in the WAL, and replay finds them. Measured: 500 acknowledged inserts, kill -9, restart, all 500 present. What it exposes is an OS crash or power loss, which can lose whatever the kernel had not flushed. Writes are roughly 15-40x faster.
off (POWDB_SYNC_MODE=off, WalSyncMode::Off) writes no WAL at all, so there is nothing to replay and the loss window is not bounded by anything: every row written since the last graceful close is gone after any unclean exit. Measured on the same setup: 500 acknowledged inserts, kill -9, restart, count = 0. The table definition survived, the rows did not. The mode exists so the benchmark harness can compare against SQLite :memory:. Never point it at data you intend to keep.

The embedded Engine checkpoints and truncates the WAL on its own once the durable log passes 64 MiB (Catalog::set_wal_checkpoint_bytes, 0 to opt out). powdb-server and powdb-cli do not: both install a WAL archive hook so retained replication history is never truncated behind a replica's back, and the automatic checkpoint never runs behind such a hook. For those two, only a graceful shutdown (SIGINT/SIGTERM) checkpoints and truncates, and during a run the log grows monotonically: 168 KB after 2,000 single-row inserts, back to 8 bytes once SIGTERM has been handled. Size the volume for the write burst between restarts, not for the size of the data.

PowQL

PowQL reads left to right. You name the table, apply operations, and project fields -- all in one pipeline.

# Define a schema (auto-increment id, a default, a required field)
type User {
  unique auto id: int,
  required name: str,
  required email: str,
  status: str default "active",
  age: int,
  city: str
}

# Insert (single row)
insert User { name := "Alice", email := "alice@example.com", age := 30 }

# Insert many rows in one statement (one fsync, one round trip, all-or-nothing).
# Keep the whole statement on one line in the CLI REPL, which buffers input
# across lines only while braces/parens stay open.
insert User { name := "Bob", email := "bob@example.com", age := 22 }, { name := "Carol", email := "carol@example.com", age := 41 }

# returning: get the affected rows back in the same statement (here the auto id)
insert User { name := "Dave", email := "dave@example.com", age := 33 } returning

# Query pipeline: source -> filter -> order -> limit -> projection
User filter .age > 25 order .age desc limit 10 { .name, .age }

# Aggregates
count(User filter .age > 25)
sum(User { .age })
avg(User filter .city = "NYC" { .age })

# Joins
User as u inner join Team as t on u.team_id = t.id { u.name, team_name: t.name }

# GROUP BY + HAVING
User group .city having avg(.age) > 30 { .city, avg_age: avg(.age) }

# JSON paths can be filtered, grouped, aggregated, ordered, and indexed
Post group .data->category { .data->category, total: sum(.data->amount) }
alter Post add index (.data->published_at)

# Subqueries
User filter .id in (Order filter .total > 100 { .user_id })

# Set operations
User filter .age > 30 union User filter .city = "NYC"

# Mutations
User filter .age < 18 delete
User filter .id = 1 update { age := 31 }
User filter .id = 1 update { age := 32 } returning   # post-update rows back

# DDL
alter User add column score: int
alter User drop column score
alter User add index .email
drop User

Run

Embedded (CLI / REPL)

powdb-cli
# or from source:
cargo run --release -p powdb-cli

Opens an interactive REPL with tab completion, command history, and meta-commands (.tables, .schema, .timing, .help). Data is stored in ./powdb_data/ by default.

Server mode

powdb-server --port 5433 --data-dir ./powdb_data
# or from source:
cargo run --release -p powdb-server -- --port 5433 --data-dir ./powdb_data

Listens on TCP with a binary wire protocol. Connect via the CLI:

powdb-cli --remote localhost:5433

Or the TypeScript client:

import { Client } from "@zvndev/powdb-client";

const client = await Client.connect({ host: "localhost", port: 5433 });
const result = await client.query("User filter .age > 25 { .name, .age }");
if (result.kind === "rows") console.table(result.rows);

Read-only snapshot serving

Serve a quiescent data directory (a restored backup or a checkpointed replica) read-only, with no write gate at all:

powdb-server --readonly --data-dir ./serve/current --port 5433

Reads are served; every mutation returns a terminal read-only error. N read-only processes can serve the same directory concurrently (a read-write open refuses while live readers exist, and vice versa), and a read-only open never mutates the directory. Embedded, use Database.openReadOnly(dir) (Node) or Database::open_read_only(dir) (Rust). See Read-only snapshot serving for the backup to restore to serve flow, the swap-directory refresh pattern, and the requirement to refresh materialized views before snapshotting.

Environment variables

The table below is powdb-server's. powdb-cli reads five of its own, listed under CLI environment variables.

Variable	Default	Description
`POWDB_PORT`	`5433`	TCP port for the server
`POWDB_BIND`	`127.0.0.1`	Interface to bind; set `0.0.0.0` behind an IPv4 platform proxy (Railway, Docker, ECS). On Fly.io use `[::]` instead, because its `.internal` network and `fly proxy` route over IPv6, so `0.0.0.0` makes the proxy reset the connection
`POWDB_DATA`	`./powdb_data`	Data directory (heap files, WAL, catalog, indexes)
`POWDB_PASSWORD`	(none)	Shared password required on connect when no named users are defined (set as env var)
`POWDB_ADMIN_USER` / `POWDB_ADMIN_PASSWORD`	(none)	Bootstrap an `admin` user on startup when both are set and that user does not yet exist (password never logged)
`POWDB_TLS_CERT` / `POWDB_TLS_KEY`	(none)	Paths to PEM cert + key; when both are set the server serves TLS
`POWDB_REQUIRE_TLS`	(off)	When set (`1`/`true`), refuse to start if a password is configured without TLS
`POWDB_IDLE_TIMEOUT`	`300`	Seconds before an idle connection is closed
`POWDB_QUERY_TIMEOUT`	`30`	Per-query deadline in seconds; cooperative cancellation stops supported scan, join, group, and mutation-discovery work and releases server admission promptly
`POWDB_QUERY_MEMORY_LIMIT`	`268435456`	Per-query memory budget in bytes (256 MiB); over-budget queries error instead of OOM-killing the server. Read by `powdb-server` only. The embedded CLI and the `powdb` crate ignore it; embedded callers set the budget in code (`Database::open_with_memory_limit` in the `powdb` facade, `Engine::with_memory_limit` in `powdb-query`)
`POWDB_TX_WAIT_TIMEOUT_MS`	`5000`	Max milliseconds a `begin` waits for a concurrent explicit transaction before failing with a timeout error instead of queueing indefinitely
`POWDB_TX_MAX_LIFETIME_MS`	`300000`	Max milliseconds one connection may hold an open explicit transaction. An explicit transaction holds the single write-admission gate for its whole lifetime, so the server bounds that lifetime: past it, the transaction is rolled back (uncommitted writes are lost), the gate is released, and the connection gets a class-3 `timeout` error and is closed. If the budget expires while a reply is being written and the write cannot finish, the connection is closed without the error frame, so treat an unexpected close during an open transaction as a timeout. The clock starts at `begin` and nothing the client sends extends it. This applies to legitimate long transactions too, so raise it (e.g. `3600000`) for servers running hour-long migrations, and split bulk loads that will not fit. `0` disables the bound and restores the pre-0.22 behavior in which one client can hold the gate, and therefore block every other connection including readers, for as long as it likes. Reaps are counted at `powdb_tx_reaped_total`
`POWDB_DB_NAME`	(accept any)	When set, the single database name this server serves; a CONNECT that explicitly names a different database is rejected
`POWDB_MAX_NESTED_LOOP_PAIRS`	`6400000`	Fallback nested-loop join candidate-pair cap; a pure non-equi join whose estimated pair count exceeds it fails before execution
`POWDB_DIRTY_PAGE_BUDGET`	`268435456`	Ceiling in bytes (256 MiB) on unflushed heap pages held in memory, shared across every table. Inside an explicit transaction those pages cannot be spilled without breaking `rollback`, so a transaction that exceeds the budget is refused with a typed error instead of growing until the process is OOM-killed. Raise it for very large bulk-load transactions, lower it on memory-capped hosts
`POWDB_SOCKET`	(off)	Path for an additional Unix-domain-socket listener served alongside the TCP listener
`POWDB_SYNC_MODE`	`full`	WAL durability: `full` (fsync before ack, fully durable) \| `normal` (bounded loss window on OS crash/power loss only, ~15-40x faster writes) \| `off` (no durability, bench-only)
`POWDB_METRICS_ADDR`	(off)	When set to `host:port` (e.g. `127.0.0.1:9090`), serve a Prometheus `/metrics` endpoint on a separate listener (metric reference). Unauthenticated: bind it to localhost or a private network, never the public internet
`POWDB_READONLY`	(off)	When set (`1`/`true`), serve the data directory read-only (snapshot serving); mutations are refused. Same as `--readonly`. See Read-only snapshot serving
`POWDB_MAX_CONNECTIONS`	`1024`	Ceiling on concurrent connections. Same as `--max-connections`
`POWDB_SHUTDOWN_TIMEOUT`	`30`	Seconds a graceful shutdown waits for connections to drain before exiting non-zero. Same as `--shutdown-timeout`
`POWDB_WAL_CHECKPOINT_BYTES`	`67108864`	WAL size in bytes at which a finished statement checkpoints and truncates the log; `0` disables it. A plain byte count, no unit suffix. Same as `--wal-checkpoint-bytes`. Has no effect on `powdb-server` or `powdb-cli` today: both install a WAL archive hook so retained replication history is never truncated behind a replica, and the automatic checkpoint never runs behind such a hook. It applies to an embedded `Engine` opened without one
`POWDB_PORT_FILE`	(off)	Path the server writes the bound listener ports to once it is listening, as `port=<n>` (plus `metrics=<n>` when the metrics endpoint is on). Written atomically before the ready log line, so a reader never sees a partial file. Pair it with `--port 0` to run a server on a free port in tests and scripts. Same as `--port-file`
`NO_COLOR`	(unset)	When set, disables ANSI colour in the log. Colour is off automatically when stdout is not a terminal
`RUST_LOG`	`info`	Log level (`debug`, `trace` for per-query timings)

Every POWDB_* value above goes through the same validator as its command-line flag. A value that does not parse refuses startup and names the variable (invalid value for POWDB_MAX_CONNECTIONS: "abc", exit 2); it is never silently defaulted. The refusal names the unit the setting is actually in, so a connection ceiling is not described as a byte count.

Two behaviours worth knowing beside the table:

A peer past POWDB_MAX_CONNECTIONS is not refused; it waits. Its TCP connection is established and then parks, unserved, until a slot frees. No bytes are read from it and the 10-second pre-auth deadline does not start until it gets one. Clients should rely on their own connect timeout rather than expecting a refusal.
SIGHUP reloads auth.json and nothing else, so a password rotated with powdb-cli passwd takes effect and a deleted user stops being able to log in without a restart. It does not reload TLS material, POWDB_PASSWORD, or any other setting. SIGTERM and SIGINT drain: the server stops accepting, closes each connection between frames with a server shutting down error, and waits up to POWDB_SHUTDOWN_TIMEOUT. A statement already executing is not interrupted.

CLI environment variables

Variable	Default	Description
`POWDB_PASSWORD`	(unset)	Password `powdb-cli --remote` authenticates with. Prefer `--password-stdin`, since `--password` is visible in `ps`
`POWDB_TLS`	(off)	When set (`1`/`true`), connect with TLS. Same as `--tls`
`POWDB_TLS_CA`	(unset)	Path to a PEM CA bundle used to verify the server certificate. Same as `--tls-ca`
`POWDB_TLS_SERVER_NAME`	(the host)	SNI/verification name to use instead of the connect host. Same as `--tls-server-name`
`POWDB_NEW_PASSWORD`	(unset)	Read by the offline `useradd` and `passwd` subcommands so a password is never typed on the command line

powdb-cli --remote accepts a Unix-domain socket path as well as host:port: an argument containing a path separator, starting with ~, or ending .sock is read as a socket. TLS over a socket is refused, because it is local and same-host. @zvndev/powdb-client speaks sockets too, through { path } instead of { host, port }.

Production checklist

Before exposing powdb-server beyond 127.0.0.1:

For a self-hostable starting point, see examples/deploy/fly.toml.

Features

Storage engine

Slotted-page heap with 4KB pages
B+tree indexes with crash-safe persistence (BIDX binary format), including scalar JSON-path indexes
Write-ahead log with statement-boundary group commit
Crash recovery (WAL replay + page-zero recovery + index rebuild)
Memory-mapped reads (zero-syscall scan path)
Compiled integer predicates (branch-free filter at the byte level)
Thread-safe concurrent reads via pread(2)/pwrite(2), with shared server admission for autocommit reads
Backup & restore: full + incremental + coarse point-in-time recovery (offline; powdb-cli backup / restore, see docs/backup-and-restore.md)

Query engine

PowQL parser + planner + executor with plan cache (FNV-1a hashing, literal substitution)
SQL frontend: a supported subset of SQL lowered to the PowQL AST, including -> / ->> JSON paths and shared plan caching (docs/SQL.md)
Joins (hash join with compound-ON residuals, plus bounded nested-loop fallback)
Nested projections (PowQL-only): one row per parent with correlated children as a native JSON array, per-parent order/limit, multi-level nesting (docs/POWQL.md)
Entity links (PowQL-only): declare a relationship once (link Order.user -> User on user_id = id) then traverse it by name: scalar o.user.name, or a labeled block orders: u.orders { ... } (the block needs a field label, which names the JSON array it returns). A scalar hop through a non-unique key is a hard error, never a silent fan-out (docs/POWQL.md)
GROUP BY, HAVING, DISTINCT
UNION / UNION ALL
Subqueries (IN, EXISTS)
Expressions in projections, filters, group keys, aggregate arguments, and order keys (arithmetic, JSON paths, string ops, BETWEEN, LIKE, IN-list)
COUNT, SUM, AVG, MIN, MAX, COUNT DISTINCT, with symmetric PowQL join semantics and explicit raw opt-out
ORDER BY (multi-expression), LIMIT, OFFSET
Window functions (ROW_NUMBER, RANK, DENSE_RANK, SUM/AVG/MIN/MAX OVER)
CAST, CASE/WHEN, COALESCE (??)
Scalar functions: UPPER, LOWER, LENGTH, TRIM, SUBSTRING, CONCAT, ABS, ROUND, CEIL, FLOOR, SQRT, POW, NOW, EXTRACT, DATE_ADD, DATE_DIFF
Materialized views with automatic dirty tracking
UPSERT with ON CONFLICT
returning on insert / update / delete (affected rows back in one round trip)
default column values and auto (auto-increment) integer columns
Prepared queries with literal substitution
EXPLAIN for query plan inspection

DDL

type (create table), drop (drop table)
alter <T> add column, alter <T> drop column (with full heap rewrite)
alter <T> add index / add unique for stored columns and scalar JSON paths; drop index for JSON-path (expression) indexes only

Server

Tokio async TCP with Arc<RwLock<Engine>> for parallel readers
Binary wire protocol (length-prefixed framing), with opt-in native typed rows/scalars for exact Bytes and PJ1 JSON
Cooperative query deadlines and disconnect cancellation
TLS support for encrypted connections
Authentication: shared password (POWDB_PASSWORD) or named users with roles (argon2id-hashed)

Pure Rust core

No libsqlite3-sys, no bindgen, no embedded C SQL engine in any published crate (the SQL frontend is pure-Rust and lowers to PowQL). The repo-root cargo build also compiles the powdb-compare benchmark harness, which does bundle SQLite from C; it is publish = false and ships nowhere
Storage, query, auth, backup, sync, and the powdb embedded facade are 100% Rust with no C dependency
powdb-server and powdb-cli pull aws-lc-sys (a C library needing cmake) through tokio-rustls. There is no tls feature to disable and no C-free build of those two today
Single cargo install on Linux and macOS (see Platform support; Windows does not compile)

Architecture

crates/
  storage/   Heap files, B+tree, WAL, catalog, page cache, row encoding
  query/     Lexer, parser, planner, executor (Engine), plan cache
  powdb/     Embedded facade crate: the engine in-process, no server
  sync/      Retained replication-unit substrate (experimental, opt-in)
  auth/      User store, roles, argon2id password hashing
  backup/    Offline backup/restore (full, incremental, PITR)
  server/    Tokio TCP server + binary wire protocol
  cli/       Interactive REPL (embedded + remote modes)
  bench/     Criterion benchmarks + regression gate (publish = false)
  compare/   PowDB vs SQLite wide-bench harness (publish = false)
  oracle/    Differential correctness oracle: runs the same fixture and query
             through PowQL, the SQL frontend, and SQLite, and compares full
             result sets (publish = false)

The engine is powdb_query::executor::Engine. It owns a Catalog (which owns Tables, each backed by a HeapFile + optional BTree indexes) and a Wal. The server wraps it in Arc<RwLock<Engine>> for concurrent access.

Benchmarks

PowDB has a benchmark regression gate that compares every workload against checked-in baselines. Run it locally before and after touching a hot path:

cargo bench -p powdb-bench              # criterion suite: 24 benchmarks (22 gated workloads), ~5 min of
                                        # measurement, plus compile on a cold target
cargo run --release -p powdb-bench --bin compare   # regression gate

The gate also runs on-demand in CI via workflow_dispatch (.github/workflows/bench.yml). It is not a required PR gate, because shared-runner noise makes it unreliable as a blocking check. The required PR gates live in ci.yml.

Run the PowDB vs SQLite comparison bench:

cargo run --release -p powdb-compare    # prints table + writes results.csv

Tests

cargo test --workspace

License

MIT License. See LICENSE for details.

Name		Name	Last commit message	Last commit date
.cargo		.cargo
.githooks		.githooks
.github		.github
bindings/node		bindings/node
clients		clients
crates		crates
docs		docs
examples		examples
scripts		scripts
site		site
.dockerignore		.dockerignore
.gitignore		.gitignore
.gitleaks.toml		.gitleaks.toml
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
Cargo.lock		Cargo.lock
Cargo.toml		Cargo.toml
Dockerfile		Dockerfile
LICENSE		LICENSE
README.md		README.md
RELEASES.md		RELEASES.md
SECURITY.md		SECURITY.md
bacon.toml		bacon.toml
deny.toml		deny.toml
docker-compose.yml		docker-compose.yml
rustfmt.toml		rustfmt.toml

Navigation Menu

Repository files navigation

PowDB

What PowDB is for

How it works

Install

Embedded in Rust

Platform support

Apple silicon / arm64

Benchmark: PowDB vs SQLite (100K rows, Apple M5 Max laptop)

Write throughput & durability

PowQL

Run

Embedded (CLI / REPL)

Server mode

Read-only snapshot serving

Environment variables

CLI environment variables

Production checklist

Features

Architecture

Benchmarks

Tests

License

About

Releases

Packages

Used by

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

PowDB

What PowDB is for

How it works

Install

Embedded in Rust

Platform support

Apple silicon / arm64

Benchmark: PowDB vs SQLite (100K rows, Apple M5 Max laptop)

Write throughput & durability

PowQL

Run

Embedded (CLI / REPL)

Server mode

Read-only snapshot serving

Environment variables

CLI environment variables

Production checklist

Features

Architecture

Benchmarks

Tests

License

About

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages