V4 of Radix, a Golang Redis Driver

by Brian Picciano - (12 min read)

Radix is a Go driver for the Redis database. The current stable release is v3, the docs for which can be found here. Over the past year (perhaps longer) I’ve been working on a new version, v4, with the aim of addressing some of the shortcomings of v3 and distilling the API a bit better.

At this point v4 is in beta. While there’s still some internal bugs and QoL improvements which need to be made, the API is roughly stable and I wouldn’t discourage anyone from using it for a non-critical project. In the coming months I intend on finishing the polish and tagging a v4.0.0 release, but in the meantime let’s go over the major changes and improvements in v4!

You can see the v4 documentation here, if you’d like to follow along with any of the particulars, and you can see the full CHANGELOG here.

Shoutouts

Before continuing I want to give to give a huge shoutout to nussjustin. Since before v3 was even stable Justin has been contributing to radix in every way possible, from running benchmarks and making very low-level performance improvements to building whole user-facing features and responding to github issues when I get lost in the woods. Thank you Justin!

RESP3

Starting at the lowest level, v4 supports new redis’s new wire protocol, RESP3. This new protocol is (mostly) backwards compatible with the previous wire protocol, and is really more an extension than anything. The new resp3 sub-package is capable of marshaling and unmarshaling all new wire types, including the streamed aggregates and streamed strings.

A major improvement made on the API level is addition of the resp.Opts type, which is used to propagate things like byte buffers and buffered readers. Doing this allows the resp3 package to reduce memory allocations without relying on something like sync.Pool, which introduces locking overhead.

There’s still some question to be answered regarding the best way for the main radix package to deal with the new push and attribute types, but the resp3 package is general-purpose enough to handle most strategies in the future.

In fact, the RESP3 protocol as a whole (and therefore v4’s associated resp3 sub-package) is totally usable outside of redis. If you’re looking for a human-readable, binary safe, fast, and simple wire protocol which already has great tooling and libraries across multiple programming languages, I highly recommend checking out RESP3.

Conn

Arguably one of the biggest design warts of v3, in my eyes, is the CmdAction type. This type required to allow for pipelining, which is a feature of redis where you can write new commands to a redis connection prior to previous ones returning their results. The major upside of pipelining is that N pipelined commands will only result in 2 system calls (a network write then a network read), rather than 2N system calls (N writes and N reads) if each command was performed independently.

The normal v3 Action type is fairly opaque, and would perform both the write and read internally without exposing any way to do some other action in between (such as performing writes/reads for other commands in a pipeline). CmdAction extends Action to allow the write and read to be performed independently, and then leaves it to the Pipeline type to deal with the batching.

v4 gets rid of the need for CmdAction, while allowing even more Action types to be pipeline-able than before (e.g. EvalScript). This was done by coalescing the Encode and Decode methods on the Conn type into a single method: EncodeDecode. By doing this we allow Actions to perform the write/read steps in a way which groups the two together, but leaves it to Conn to actually perform the steps in its own way.

Because Conn now has knowledge of which read/write steps go together, it’s possible to perform pipelining in nearly all cases. Aside from using the Pipeline type manually, the v4 Conn is able to automatically pipeline most Actions when they are performed concurrently on the same Conn. v3 had a similar feature, called “implicit pipelining”, but v4 rebrands the feature as “connection sharing” since the mechanism is slightly different and the applicability is broader.

Despite the apparent simplicity of the change (combining Encode and Decode methods), this resulted in probably the largest code difference between v3 and v4, involving the most complex new logic and package-wide refactorings. But the end result is a simpler, smaller API which can be applied to more use-cases. A great win!

Pool

In v3 the connection pool, the Pool type, was implemented with the assumption that each Action (or CmdAction) would borrow a Conn for the duration of the Action. As such the Pool expects to be creating and destroying connections as load increases and decreases; if number of concurrent commands goes up then number of connections required to handle them goes up as well, and vice-versa.

Down the road the Pool became responsible for performing implicit pipelining as well. This allowed for grouping together many commands on the same connection, reducing pressure on connection creation greatly, but nevertheless the Pool kept that same general pattern of dynamic connection pool sizing.

In v4 there is no longer the assumption that each command gets its own connection, and in fact that assumption is flipped: each connection is expected to handle multiple commands concurrently in almost all cases. This means the Pool can get rid of the dynamism, and opt instead for a simple static connection pool size. There is still room in the API for some dynamic connection sizing to be implemented later, but it’s mostly unnecessary now.

Some care should be used with commands which can’t be pipelined, for example blocking commands like BRPOPLPUSH and XREAD. These commands, ideally, should be performed on an individual Conn created just for that purpose. Pool will properly handle them if needed, but with the caveat that the Action which will essentially remove a Conn from the Pool for its duration.

The new Pool is vastly simpler in implementation than the old, as most of the complexity has been moved into Conn. Really this whole section is an extension of the refactoring which was started by the changes to Conn.

MultiClient

In v3 there was a single Client type which was used to encompass Conn, Pool, Sentinel, and Cluster, with the aim that users could just use Client in their code and easily swap out the underlying implementation as needed.

In practice this didn’t work out. The original Client type only had a Do method for performing Actions, which would always perform the Actions against the primary instance in the case of Cluster and Sentinel. Cluster and Sentinel ended up being extended with DoSecondary methods, and Cluster required its own constructor for Scanner, so if you used any of those features you would not be able to use Client.

v4 improves this situation by introducing the MultiClient interface, which is implemented by both Cluster and Sentinel, while Conn and Pool only implement Client. Client is intended for clients which interact with only a single redis instance, while MultiClient is intended for use by clients which encompass multiple redis instances, and makes the distinction between primary and secondary instances.

In general, users will want to use MultiClient in their code and swap the underlying implementation as their infrastructure evolves. When using only a single Pool, one can make it into a MultiClient using the new ReplicaSet.

One can also implement their own MultiClient’s fairly easily, to handle their own custom sharding or failover systems. It’s not a common use-case, but it’s cool that existing types like Scanner will still continue to work.

Contexts

A common feature request of v3 was for support for Go’s Contexts, which would allow callers to unblock blocked operations in a dynamic way. There wasn’t a clear way to incorporate Contexts into v3 without greatly expanding the API (something the Go standard library has had to do), and so I saved them for v4.

In v4 all operations which might potentially block accept a Context argument. This takes the place of timeout options and some trace events which were used in v3, and in general simplifies things for the user.

This was a change for which there is not much to talk about, but which required a lot of work internally. Go’s Contexts do not play nicely with its networking primitives, and making this all work alongside connection sharing and pipelining is a really hairy puzzle (for which there’s a few open bugs still). I may one day write a blog post just about this topic, if I can figure out how to explain it in a way which isn’t completely mind-numbing.

Configuration

Constructors in v3 took advantage of the functional options pattern for accepting optional parameters. While this pattern looks nice, I’ve since grown out of love with it. The implementation is a lot more complex, its behavior is more ambiguous to users in certain cases (what happens if the same option is passed in twice?), it makes documentation more complex, and a slice of option functions isn’t inspectable or serializable like a struct is.

v4 uses a config struct pattern, but in a different way than I’ve generally seen it. See Pool’s constructor for an example. This pattern is functionally the same as passing the config struct as an argument to the constructor, but I think it results in a nicer grouping in the documentation.

Smaller Changes

There’s some smaller sets of changes which are worth mentioning. These didn’t result in huge, package-wide changes, but will be useful for users of specific functionality.

Action Properties

v4’s Action type has a Properties method which returns a struct containing various fields which are useful for client’s performing the Action. This is an improvement over v3’s Action, which had no such method, in that it’s more extensible going forward. Those implementing their own custom Actions should take care to understand the Action properties.

PubSub

The v4 PubSubConn has been completely redesigned from v3’s implementation. The old design tried to do too much, and resulted in weird edge-cases when trying to tear down a connection that a user would have to handle themselves. The new design is simple both in implementation and usage.

Tracing

The v4 trace sub-package has been extended to support tracing Sentinel events, but at the same time has been cleaned out of all events which could be otherwise inferred by using Context values or wrapping an interface like Conn, Action, etc…

What’s Next

Obviously the most immediate goal is to get v4 stable and tagged. Once that’s done I’m sure there will be many small bugs, feature requests, etc… which come up over time, and I’ll do my best to address those as quickly as I can. I’m very excited to start using v4 in my own day-to-day work like I currently do for v3; it has a lot of great improvements and new flexibility that will make using Go and redis together an even better experience than it already is.

That all said, I don’t expect there to be a radix v5. I have a lot of other projects I’d like to work on, and radix is a huge time-sink. As time goes on v4 will stabilize further and further, until all that’s left is for it to gain additional support for whatever new crazy features redis comes up with. My hope is that the existing API is flexibile enough to allow others to fill in those gaps without any major changes to the existing code, and radix v4 can be the final major radix version.