See the design document included in this PR for full details.

Problem Statement

The current connection pool implementation is slow to open new connections and does not follow async best practices.

Connection opening is serialized, causing delays when multiple new connections are required simultaneously. This is done using a semaphore to rate limit connection creation. When multiple new connections are requested, they queue up waiting for the semaphore. Once acquired, the thread opens the connection and releases the semaphore, allowing the next thread to proceed. This approach was initially designed to prevent overwhelming the server but can lead to significant delays, especially in high-latency environments.

Async requests are also serialized through an additional queue. When an async request is made, it is added to a queue, and a reader thread processes these requests one by one. This method was chosen due to the lack of async APIs in native SNI, resulting in synchronous handling of async requests on a dedicated thread.

Design Goals

• Enable parallel connection opening.
• Minimize thread contention and synchronization overhead.
• Follow async best practices to reduce managed threadpool pressure and enable other components of the driver to modernize their async code.

Overview

The core of the design is the Channel data structure from theSystem.Threading.Channels library (available to .NET Framework as a nuget package) (Also see Stephen Toub's introhere). Channels are thread-safe, async-first queues that fit well for the connection pooling use case.

A single channel holds the idle connections managed by the pool. A channel reader reads idle connections out of the channel to vend them to SqlConnections. A channel writer writes connections back to the channel when they are returned to the pool.

Pool maintenance operations (warmup, pruning) are handled asynchronously as Tasks.

Transaction-enlisted connections are stored in a separate dictionary data structure, in the same manner as the WaitHandleDbConnectionPool implementation.

This design is based on thePoolingDataSource class from the npgsql driver. The npgsql implemenation is proven to be reliable and performant in real production workloads.