In some circumstances users may have connection pools configured without
any SocketAddresses ready to go. This is particularly likely in service
discovery configurations. Right now, the effect of attempting to use
such a pool is two fold. First, we'll emit a bunch of error level logs
telling users we have no addresses. Second, we'll fall into the
exponential backoff phase of connection establishment.
The first property is annoying, but the second one is actively harmful.
If your construction is timed incorrectly, we'll have the awkward
problem of burning a bunch of CPU trying to create connections we know
we cannot, and then a lengthy delay after the addresses are actually
configured before we start trying to use them. That's the worst of all
worlds.
This patch adds logic to detect the attempt to create connections when
we don't have any configured addresses and delays them. This path should
improve performance and ergonomics when in this mode.
Authored-by: Cory Benfield <lukasa@apple.com>
Motivation:
Some Redis commands are very connection specific that have impacts on future access that makes it difficult in the current
checkout-use-return cycle that `RedisConnectionPool` uses.
Developers need a way to borrow a specific connection, chain several commands together, and then return the connection to the pool.
Modifications:
- Add: `leaseConnection` method to `RedisConnectionPool` which provides a connection from the pool and returns it after a provided closure's ELF resolves
- Add: `allowSubscriptions` property to `RedisConnection` for controlling the ability to make PubSub subscriptions
- Add: `RedisClientError.pubsubNotAllowed` case for when `RedisConnection.allowSubscriptions` is set to `false` and a subscription was still attempted
Result:
Developers should now have an "escape hatch" with `RedisConnectionPool` to do limited exclusive chains of operations on a specific connection.
Motivation:
Users of Redis will frequently want to be able to run queries in
parallel, while bounding the number of connections they use. They will
also often want to be able to reuse connections, without having to
arrange to manage those connections themselves. These are jobs usually
done by a Connection Pool.
This new connection pool will conform to `RedisClient` so a pool of clients and a single connection are interchangeable.
Connection Pools come in a wide range of shapes and sizes. In NIO
applications and frameworks, there are a number of questions that have
to be answered by any pool implementation:
1. Is the pool safe to share across EventLoops: that is, is its
interface thread-safe?
2. Is the pool _tied_ to an EventLoop: that is, can the pool return
connections that belong on lots of event loops, or just one?
3. If the pool is not tied to an EventLoop, is it possible to influence
its choice about what event loop it uses for a given connection?
Question 1 is straightforward: it is almost always a trivial win to
ensure that the public interface to a connection pool is thread-safe.
NIO makes it possible to do this fairly cheaply in the case when the
pool is only used on a single loop.
Question 2 is a lot harder. Pools that are not tied to a specific
EventLoop have two advantages. The first is that it is easier to bound
maximum concurrency by simply configuring the pool, instead of needing
to do math on the number of pools and the number of event loops. The
second is that non-tied pools can arrange to keep busy applications
close to this maximum concurrency regardless of how the application
spreads its load across loops.
However, pools that are tied to a specific EventLoop have advantages
too. The first is one of implementation simplicity. As they always serve
connections on a single EventLoop, they can arrange to have all of their
state on that event loop too. This avoids the need to acquire locks on
that loop, making internal state management easier and more obviously
correct without having to worry about how long locks are held for.
The second advantage is that they can be used for latency sensitive
use-cases without needing to go to the work of (3). In cases where
latency is very important, it can be valuable to ensure that any Channel
that needs a connection can get one on the same event loop as itself.
This avoids the need to thread-hop in order to communicate between the
pooled connection and the user connection, reducing the latency of
operations.
Given the simplicity and latency benefits (which we deem particularly
important for Redis use-cases), we concluded that a good initial
implementation will be a pool that has a thread-safe interface, but is
tied to a single EventLoop. This allows a compact, easy-to-verify
implementation of the pool with great low-latency performance and simple
implementation logic, that can still be accessed from any EventLoop in
cases when latency is not a concern.
Modifications:
- Add new internal `ConnectionPool` object
- Add new `RedisConnectionPool` object
- Add new `RedisConnectionPoolError` type
- Add tests for new types
Results:
Users will have access to a pooled Redis client.
Konrad `ktoso` Malawski
Fabian Fett
Joannis Orlandos
Nathan Harris
Lukasa