co-log: Composable Contravariant Combinatorial Comonadic Configurable Convenient Logging

This blog post illustrates the architecture of the co-log library: a composable Haskell logging library that explores an alternative way of logging. I’m not going to cover how to implement the ElasticSearch backend for the logging library or how to make concurrent logging fast. Instead, the blog post explains the core ideas of the new design. I’m going to describe in details and with examples how one can build a flexible, extensible and configurable logging framework using different parts of Haskell — from monad transformers and contravariant functors to comonads and type-level programming with dependent types.

If you want to go straight to the library’s source code, you can follow the link below:


There are already several logging frameworks within the Haskell ecosystem and every library has its own idea and architecture:

So you might ask, why create another library? The motivation behind co-log is to explore a new strategy that allows having a composable and combinatorial logging library that is easy to extend and use. This is achieved by decomposing the logging task into smaller pieces:

  • What to log: text, message data type, JSON
  • Where to log: to the terminal, to some file, to an external service
  • How to format the output: coloured text or JSON
  • How to log: with logger rotation, only to stderr or something else
  • What context to work in: pure or some IO
  • How to change context: append logs to in-memory storage or change filesystem

The co-log framework is currently split into two packages:

  • co-log-core: lightweight library with core data types and basic combinators
  • co-log: implementation of a logging library based on co-log-core

In the following sections, I’m going to describe the main ideas behind these two packages alongside with some implementation details.


This section introduces the fundamental piece in the co-log design: the LogAction data type.

Data type🔗

The main piece of the co-log logging library is the following structure:

This is a wrapper around a simple function. It has two type variables:

  • m: monad in which the logging is happening
  • msg: the logging message itself

LogAction specifies the type of message you want to log and the context in which you want to perform the logging. So you can tune and configure your action (or, even better: create a compound action by combining smaller pieces) and use it later.

Here is an example of a very straightforward action:

NOTE: For simplicity purposes this blog post uses String data type for logging. In practice, it’s better to use Text or ByteString data types instead of String as they provide better performance. Even better, you can use Backpack and implement a general interface around backpack-str package.

How to use it?🔗

Once you’ve created LogAction, you need to use it in your application. There are multiple ways to use LogAction in your code since it’s just a value:

  • Pass it as an argument to your functions explicitly each time
  • Use Handle design pattern and store LogAction there
  • Store it in a separate ReaderT layer
  • Store it in your own ReaderT environment
  • Use an extensible-effects library like freer-simple
  • Use caps framework
  • Etc.

Every solution has its own advantages and drawbacks. That’s why the LogAction data type is in the co-log-core package: so it’s possible to experiment with different approaches and use the main concepts of the co-log library without bringing extra dependencies to your project.

The simplest solution is to pass LogAction explicitly as an argument to every function where you need logging. The co-log-core library has this helpful combinator:

So you can pass messages to actions using this operator:

In the remaining part of the blog post I’m going to use the following approach supported by co-log.

First, let’s introduce a newtype wrapper around ReaderT that stores LogAction in its own environment:

This data type looks scary but the idea is simple: it just stores LogAction m msg in the ReaderT environment.

The co-log-core package also has its own lens-like HasLog typeclass that allows to get and set LogAction in your environment (so you can just add LogAction to your ReaderT context instead of using LoggerT monad transformer):

Now we can write a function that logs messages using LogAction from the context:

We now need some way to execute actions with logging:

After implementing usingLoggerT we can now play with our logging framework.

And the output is exactly what we expected:

Starting application...
Finishing application...



Let’s see how to make our LogAction composable. Currently, we have only one action that prints String to stdout. What if we also want to print the same String to stderr in addition? Or to some file? Of course, we can create a separate LogAction that does just that:

We start to notice a pattern: we often want to perform multiple actions over a single message. This is where Semigroup comes in:

So instead of manually specifying what we want to do on every message, we can create our LogAction by combining different smaller and independent pieces.


The Monoid instance is even simpler:

It adds the empty LogAction that does nothing to the Semigroup instance. It’s quite straightforward to show that the above instances satisfy associativity and neutrality laws. The nice thing about the Monoid instance is the ability to disable logging. Usually, logging libraries provide some extra transformer with a name like NoLoggingT which you can put on top of your monad transformer tower to disable logging. But with a Monoid instance for LogAction you can just pass mempty as your logging action and that’s all. This approach also works quite well if you want to disable logging only in some specific piece of your code.

Contravariant functors🔗

This section covers the contravariant family of typeclasses regarding LogAction. The typeclasses themselves are implemented in the contravariant package. This section assumes some basic knowledge for Contravariant functors, if you’re not familiar with them, I strongly recommend watching the following talk by George Wilson:



Let’s first have a look at Contravariant typeclass and its instance for LogAction:

Here is how you can think of this instance: if you know how to log messages of type b and you know how to convert messages of type a to type b, then you also know how to log messages of type a. You just need to convert the a message to b and pass it to your action.

Here is an example of how it can be useful. Let’s say that instead of just logging String, you want to log a more complex Message data type:

But you only have LogAction that prints String. This is no longer a problem! You can write the formatting function:

And then you can use contramap to log messages instead of String.

The output is the following:

[Debug] Starting application...
[Info] Finishing application...

The problem of showing the output is handled separately from the problem of providing the input to the loggers. If you want to format the output in a different way (as JSON for example) you can just switch the formatting function to a different one.


Okay, now we would like to discard any Debug messages from the output. It’s easy to do after writing a contravariant filter function:

Now you can filter out all the Debug messages in the following way:


There’s even more! We’re not satisfied with printing only Severity of the message. We would like to print the timestamp of the logging message as well. This is very useful for further log analysis. But with contramap we can’t do that because taking the current time is an impure function. So let’s implement cmapM function and see how it helps:

Look at how similar it is to contramap:

Now we can extend our Message data to RichMesssage that also stores UTCTime in it.

After writing the formatting function for RichMessage we can enjoy an automatically appended timestamp to every message!

And now we have a more verbose output:

[Info] [11:54:31.809 13 Sep 2018 UTC] Finishing application...


The Functor-Applicative-Alternative family of typeclasses is well-known to most Haskellers. But there are similar typeclasses for contravariant data types. Specifically, Contravariant-Divisible-Decidable. Let’s look at the Divisible typeclass definition and its instance for LogAction:

What this instance means is that if you know how to split some complex data structure into smaller pieces and if you know how to log each piece independently, you can now log the whole data structure.


Decidable is similar to Alternative but for the Contravariant family of functors.

The meaning of the Decidable instance for LogAction is that you can look at your logging message and decide what exactly you want to log. And if you know how to log every decision independently, you then can log the message.


Using the above instances and combinators from the co-log library, we can now have a combinatorial logging library.

Let’s first introduce the data types we want to log:

We also need couple helper functions in order to use the contravariant combinators:

Then we’re going to introduce some basic logging actions:

And then, we can add combinators:

So we can log our Car data type in the following way:

And the output is:

Logging make...
Finished logging make...
Logging model..
Finished logging model...
Logging pistons...

This example might look not so convincing. But the approach becomes more useful when you want to log sophisticated data structures as it allows you to divide the message into smaller pieces and decide what to log.


Now, let’s talk about one of the most interesting parts of the co-log library. There is the Comonad typeclass which is dual to Monad. It’s implemented in the comonad package:

And there’s a Comonad instance for the Traced data type:

If you look closely at Traced data type, you might recognise similarities with LogAction. Indeed, LogAction can be defined as a special case of the Traced data type.

Since LogAction is just a special case of the Traced comonad, we can implement the extract and extend functions for LogAction.

Unfortunately, we can’t implement the Comonad instance for LogAction due to the interface mismatch. It would be possible only if LogAction was defined as a specialized version of the Traced comonad. However, by doing so we would lose other useful properties of defining LogAction as a separate newtype. However, we can still use the comonadic aspect of the LogAction structure. Here is a couple of simple examples of LogAction comonad usage:

What this comonadic aspect of the interface means is that you can extend the existing LogAction by passing an additional payload to every message. This only works if your message is a Semigroup (or Monoid) and preferably a commutative Semigroup (or Monoid). For example, Map String String. If you log key-value pairs, you might want to add additional entries depending on the local context. Turns out this is extremely useful for structured logging: if you log JSON values, you might want to add extra keys or tags or values for specific functions and here is where the comonadic interface becomes helpful.

Configurable & Convenient🔗

This section covers how dependent map, type-level programming and the -XOverloadedLabels language extension are used to implement a flexible and extensible interface for logging configuration.

Earlier I’ve introduced the RichMessage data type that allows to extend a simple Message with the result of an IO action:

Currently, it stores only the timestamp of the logged message. But in real life we might want to display more information on every logging message:

  • Id of the current thread
  • Id of the current process
  • File size of the file where we currently write logs
  • Application memory usage
  • Number of currently active users (why not?)
  • Any other thing you can imagine

If we only have a simple plain Haskell record we can’t easily extend it. Of course, you can always implement your own RichMessage and format in any way you want, but usually developers expect at least some configuration capabilities from a logging library. And configuring your output through boolean flags is not that convenient. That’s why co-log provides an extensible record interface to make it easier to add or remove some fields and I’m going to describe the implementation below.

The idea behind this solution is the following: let’s use some map to store the actions that extract the required information. So instead of specifying control options for some predefined set of possible message fields, we can just modify the map by adding or removing actions we want to execute on every call to the logging function. But since the actions might return values of different types, we need to use some dependent map.

The assumption here is that usually we configure logging for our application only once at the start of our application and then we only query those actions. So we need some map where construction and modification operations should be supported but their performance is not that important. The efficiency of the lookup function, on the other hand, should be quite high. Fortunately, there’s a structure that implements the required interface. The data structure itself is implemented in the typerep-map package. If you want to learn about the implementation details of the typerep-map library, you can read the following blog post:

We can build extensible records on top of the typerep-map, and I’m going to explain how it’s done in co-log.

First, let’s introduce a type family that maps arbitrary string tags to types.

The type family is open, so users can extend it with anything they want.

Now let’s introduce a data type that stores values of type FieldType inside some monad.

We can now parametrize TypeRepMap with MessageField:

So FieldMap is a mapping from type level string to runtime value inside some monad m, where the type of the runtime value is defined by the value of the type-level string.

There is an existential wrapper in the typerep-map library that allows to create TypeRepMap from a list due to IsList instance:

Now we can extract monadic actions that return us values by doing lookup @”threadId”. Since it’s just a map, you can configure the output by putting additional actions in the map with the insert function or by removing some actions with the delete function. You only need to implement the formatting function for your FieldMap or use the default one from the co-log library.

But that’s not all. There’s the OverloadedLabels extension since GHC 8.0.1 that allows us to specify such FieldMaps in a more convenient way. After writing the following instance:

The OveloadedLabels extension allows to write #foo instead of fromLabel @”foo”. So, by having the above instance for the function we can pass arguments to labels. We’re now able to define defaultFieldMap in a more concise way:

Here is how the output looks like with ThreadId and without it, for example from the library’s playground:

Logging example
Logging example

The solution described above is currently implemented only for the IO part of the message extension. But there are plans to make the Message data type extensible as well (see issue below):

Conclusion & unexplored opportunities🔗

The LogAction data type is very simple on its own, but together with different instances, it brings a lot of power:

  • Semigroup: perform multiple actions for the same message
  • Monoid: the action that does nothing
  • Contravariant: the ability to consume action of the different type
  • Divisible: log both types of messages if you know how to log all of them
  • Decidable: decide what to log depending on the message
  • Comonad: add an extra monoidal payload to message possibly using context

There’s also the ComonadApply typeclass which purpose in application to LogAction is yet to be explored. And it would be really great to implement different logging backends for the co-log library (ElasticSearch, PostgreSQL, etc.).


If you want to play with this approach in PureScript, there is a library that implements a similar idea: