Download the PHP package kosadchiy/laravel-parallel-db without Composer

Laravel Parallel DB Queries

Parallel SQL queries for Laravel using native async database APIs:

• PostgreSQL: ext-pgsql (pg_send_query_params, pg_socket, pg_consume_input, pg_connection_busy, pg_get_result)
• MySQL: ext-mysqli (MYSQLI_ASYNC, mysqli_poll, reap_async_query)

No child processes, no queues, no ReactPHP, Amp, or Swoole.

Status: beta. The package is usable, but APIs and behavior may still evolve and are still being tested in real-world scenarios.

Why

When one request needs multiple independent SQL calls, sequential execution adds latency. This package runs them in parallel on separate DB connections in the same PHP process.

Typical use cases:

• building dashboards that need several expensive aggregate queries for different widgets;
• assembling analytics pages that combine totals, trends, leaderboards, and recent activity;
• fetching independent datasets for a single API response, such as users, transactions, invoices, and alerts;
• loading admin or reporting screens where multiple tables, counters, and summaries can be queried independently;
• reducing end-to-end latency in read-heavy endpoints that would otherwise wait on one slow query after another.

This package is most useful when queries are independent, read-heavy, and latency-bound. For tiny queries or cold-start paths, the overhead of extra connections, polling, and result collection can outweigh the benefit and even make the request slower. It is also a poor fit for strongly dependent query chains or workloads that require a shared transaction.

High-level design

Core components

• ParallelDatabaseManager - Laravel-facing entrypoint, accepts mixed query inputs.
• QueryCompiler - compiles builder payloads into CompiledQuery.
• ParallelExecutor - bounded scheduler + timeout + result collection.
• DriverRegistry - maps DB driver names to async implementations.
• PostgresAsyncDriver, MySqlAsyncDriver - native async transport.
• ConnectionConfigResolver + *ConnectionFactory - per-query dedicated connection creation.
• ConnectionPool - optional process-local reuse of idle async connections between runs.

Execution flow

1. User passes map of queries (Builder, Eloquent, or Closure returning one of them).
2. Compiler normalizes all inputs to CompiledQuery.
3. Executor keeps pending/running sets and starts up to maxConcurrency queries on dedicated connections.
4. Drivers poll readiness (stream_select() for PostgreSQL sockets, mysqli_poll() for MySQL links).
5. Ready queries are reaped and transformed into QueryResult.
6. Completed connections are returned to the pool when pooling is enabled.
7. Executor continues until all queries are finished or timeout is reached.

Laravel versions

• Laravel 12.x
• PHP 8.3+

Installation

Optional extensions by driver:

• ext-pgsql
• ext-mysqli

Publish config:

Configuration

• default_max_concurrency - default per-run concurrency limit used by DB::parallel().
• default_timeout_ms - default batch timeout in milliseconds for the whole parallel run.
• error_mode - whether to stop on the first failure (fail_fast) or return failed entries alongside successful ones (collect).
• pool.enabled - enables process-local reuse of idle async connections between runs.
• pool.max_idle_per_key - caps how many idle pooled connections are retained per connection config.
• drivers.pgsql.enabled - enables the PostgreSQL async driver.
• drivers.mysql.enabled - enables the MySQL async driver.

Usage

Default DB manager path

Accessing collections and models

Using closures

With options

Via specific connection

Result format

Each input key maps to QueryResult:

• sql
• bindings
• type
• rows
• rowCount
• lastInsertId
• durationMs
• connectionDriver
• success
• error

Connection pooling

Idle async connections are pooled by default and reused across run() calls inside the same PHP process.

Connection pooling is especially useful in long-running environments such as Laravel Octane, where worker processes can reuse async connections across many requests. Keep in mind that the pool is process-local: each worker may retain up to max_idle_per_key idle connections per connection config, so total open connections should be sized with your worker count and database limits in mind.

For long-running workers and higher connection counts, consider a database-side pooler such as PgBouncer for PostgreSQL or ProxySQL for MySQL.

• pool key is based on driver/config;
• one pooled connection can still serve only one active query at a time;
• hot runs should show lower connection overhead;
• this is process-local reuse, not a shared external pool.
• maxConcurrency limits concurrently running queries per run();
• max_idle_per_key limits how many idle pooled connections are retained for one connection config;
• if pooling is enabled, a long-lived process can keep more open connections than maxConcurrency, because completed connections may stay idle in the pool for reuse.

Error behavior

• fail_fast: first failed query stops the batch and throws ParallelQueryFailedException.
• collect: executor continues and returns failed entries with success=false.

Write-query caveats

Writes are allowed (INSERT/UPDATE/DELETE/DDL), but:

• every query uses a separate connection;
• no shared transaction across the batch;
• global rollback is impossible;
• commit ordering across queries is not guaranteed.

Use parallel writes only when this behavior is acceptable.

Driver notes

PostgreSQL

• Uses parameterized async calls (pg_send_query_params) with ? -> $1..$N conversion.
• Wait uses socket readiness (pg_socket + stream_select).

MySQL

• Uses MYSQLI_ASYNC + mysqli_poll + reap_async_query.
• The MySQL async path does not use prepared statements and interpolates bindings with mysqli::real_escape_string, which is a transport-level limitation of this approach.
• Bindings are interpolated with mysqli::real_escape_string and strict typing rules.

Limitations and controversial points

1. Async API parity differs between ext-pgsql and ext-mysqli.
2. The MySQL async path does not use prepared statements; bindings are interpolated with escaping as a transport-level limitation of this approach.
3. stream_select() has platform limits and FD-scaling caveats for very large sets.
4. Core results are returned as raw rows, but QueryResult provides explicit toCollection() and toEloquentCollection() helpers for post-processing.
5. Parallel writes are independent operations, not an atomic group transaction.

Benchmark

Synthetic PostgreSQL benchmark setup:

• 7 independent queries per run;
• 5 queries with configurable pg_sleep(...) delay;
• 2 regular SELECT queries;
• 2 warmup runs, then 10 measured iterations;
• sequential and parallel paths consume the same result sets.

This benchmark is intentionally synthetic and latency-bound. It is useful for showing when overlapping independent query latency helps, but it should not be read as a universal production speedup claim. Real-world gains depend on query cost, result size, connection overhead, database load, pooling behavior, and maxConcurrency.

Pool settings also affect warm-run results. With the default config, up to 3 idle connections per connection key are retained between runs.

Observed trends from the local test environment:

• higher maxConcurrency is not always better; the best value depends on the workload;
• longer per-query latency increased the benefit of parallel execution;
• for tiny queries, connection setup, polling, and result collection overhead could outweigh the gains.

Suggested benchmark matrix for comparing different latency profiles and scheduler settings:

Cold-start overhead looks different and should be measured separately. In the following run, pooling was disabled and warmup=0, so the async path had to pay full connection setup cost. For this kind of cold tiny-query scenario, p50 is more representative than avg, because a single startup outlier can skew the sequential mean.

Cold-start threshold sweep for the same environment (pool=off, warmup=0, maxConcurrency=3):

Takeaway from these synthetic benchmarks:

• warm runs benefited earlier, because pooled connections reduced setup overhead;
• cold runs with pool=off showed that parallel execution was not worthwhile for tiny queries;
• in this environment, the cold-path break-even point appeared somewhere between 20ms and 50ms of per-query latency;
• the optimal maxConcurrency depended on the workload: around 3 for shorter waits and higher for more latency-bound queries.

Testing

Unit tests cover:

• query compilation from Builder/Eloquent/Closure;
• bounded concurrency;
• timeout handling;
• fail-fast and collect-errors behavior;
• key-preserving result mapping.

Run:

Integration tests cover:

• public API query shapes (Query\Builder, Eloquent\Builder, Closure);
• raw compiled write queries through ParallelExecutor;
• collect vs fail_fast;
• timeout behavior;
• transaction visibility across separate async connections.

Run:

Notes:

• the integration suite uses orchestra/testbench;
• it requires real PostgreSQL and MySQL instances plus ext-pgsql and ext-mysqli;
• when those services are unavailable locally, the integration suite skips cleanly;
• in CI, integration tests run against dedicated PostgreSQL and MySQL service containers.