July 6, 2015

GravatarAndrew Thaddeus Martin

Announcing yesod-table

Over the last two years, I've seen the need for safe dynamic table-building in half a dozen yesod projects I've worked on. After several design iterations, the result of this experience is yesod-table, which saw its first stable release last week. This blog post will contain code excerpts, but you can also look at the documentation the full example app on github, which can be compiled and run.

Naive Solution

Before getting into specifics about yesod-table, I want to take a look at the naive table-building strategy and identify the common pitfalls. Let's say that you have a data types Color and Person:

data Color = Red | Green | Blue | Purple
  deriving (Show)
data Person = Person
  { firstName     :: Text
  , lastName      :: Text
  , age           :: Int
  , favoriteColor :: Color

We have a list of Person (let's call it people), and we want to show them all in a table. You could write out a hamlet file like this:

      <th>First Name</th>
      <th>Last Name</th>
    $forall p <- people
        <td>#{firstName p}
        <td>#{lastName p}
        <td>#{show (age p)}

And there it is. This is the simplest solution to building a table. In fact, if you've worked on web applications for any length of time, you've probably written code like this before. I've implemented this pattern in PHP+html, in rails+haml, and in yesod+hamlet projects. And every time, it is unsatisfactory.

Problems With Naive Solution

Let's take a look at three reasons why this solution leaves us wanting something more:

  • Duplication. After building a few tables this way, you realize that you are copying the HTML elements and the list iteration ($forall) every time.
  • Non-composability. If I want to build a similar table, one that shows the same fields but additionally has a column for favoriteColor, I have to copy the whole thing. I can't glue another piece onto the end.
  • Breakable Invariant. If we do decide to add a favoriteColor column, we might try simply adding <td>#{show (favoriteColor p)} to the end. This would cause incorrect behavior at runtime, because we would have forgotten to add <th>Favorite Color to the table header. The problem is that we have an invariant not captured by the type system: thead and the tbody loop must have the same number of <th>/<td> elements, and the order must match.

In particular, the last issue (breakable invariant) has been a source of great pains to me before. On a three-column table, you are less likely to forget the <th> or put it in the wrong place. As the table gets larger though (six or more columns), these mistakes become easier to make, and it's harder to be sure that you did the right thing until you see it at runtime.

Example with yesod-table

So let's take a look at how yesod-table addresses these issues. The module it provides should be imported as follows:

import Yesod.Table (Table)
import qualified Yesod.Table as Table

Let's build the same table we saw earlier:

peopleBasicInfoTable :: Table site Person
peopleBasicInfoTable = mempty
  <> Table.text   "First Name" firstName
  <> Table.text   "Last Name"  lastName
  <> Table.string "Age"        (show . age)

And then we can feed it data and render it with buildBootstrap:

-- Although it's called buildBootstrap, it builds a table just fine
-- if you aren't using bootstrap. It just adds bootstrap's table classes.
getExamplePeopleR = defaultLayout $ Table.buildBootstrap peopleTable people

Explanation of Internals

The key to this approach is looking at a table pattern (called a Table in this library) as a collection of columns, not a collection of rows. From the yesod-table source, we have:

newtype Table site a = Table (Seq (Column site a))
  deriving (Monoid)

data Column site a = Column
  { header :: !(WidgetT site IO ())
  , cell :: !(a -> WidgetT site IO ()) 

Each column is just the content that will go in the <th> (the value of header) and a function that, given the data for a table row, will produce the content that belongs in the <td>. A table is trivially a collection of columns and gets a Monoid instance from Seq for free (for those unfamiliar, Seq a is like [a] but with different performance characteristics). Consequently, any two Tables that are parameterized over the same types can be concatenated. As a final note of explanation, the Table.text function that we saw above just a helper for building singleton tables. So, the three Tables below are equivalant:

import qualified Data.Sequence as Seq
import qualified Data.Text as Text
-- These three generate a single-column table that displays the age.
reallyEasyToReadTable, easyToReadTable, hardToReadTable :: Table site Person
reallyEasyToReadTable = "Age" age
easyToReadTable = Table.text "Age" (Text.pack . show . age)
hardToReadTable = Table.Table $ Seq.singleton $ Table.Column 
  (toWidget $ toHtml "Age")
  (toWidget . toHtml . show . age)

As should be clear, the convenience functions for singleton Tables should always be preferred.

How Is This Better?

Now to address the most important question: Why is this better than what we had earlier? Firstly, consider the issue of the breakable invariant. This is now a non-issue. Imagine that we modified the earlier table to show a person's favorite color as well:

peopleFullInfoTable1 :: Table site Person
peopleFullInfoTable1 = mempty
  <> Table.text   "First Name"     firstName
  <> Table.text   "Last Name"      lastName
  <> Table.text   "Age"            (show . age)
  <> Table.string "Favorite Color" (show . favoriteColor)

The table is correct by construction. You cannot forget the column header because it's part of the Column data type. You're less likely to make this mistake, because now that information is directly beside the content-extracting function, but even if you somehow typed this instead, you would get a compile-time error:

  <> Table.string (show . favoriteColor)

Secondly, we can look at duplication. All the table-rendering logic is moved into buildBootstrap (and you can write your own table renderer if that one is not satisfactory). The Table that we are using now has neither the HTML elements nor the list iteration that we dealt with earlier.

Finally, we can look at composability. As an alternative way of adding the column for a person's favorite color, we could write:

peopleFullInfoTable2 :: Table site Person
peopleFullInfoTable2 = mempty
  <> peopleBasicInfoTable
  <> Table.string "Favorite Color" (show . favoriteColor)

Additionally, if we need to promote this Table to work on something like Entity Person (if it was backed by persistent), we could do this:

-- You need to use ekmett's contravariant package
peopleEntityFullInfoTable :: Table site (Entity Person)
peopleEntityFullInfoTable = contramap entityVal peopleFullInfoTable2

I won't go into contravariant functors here, but it's a very useful pattern for working with Tables. The astute reader will notice that the monoidal composition pattern shown earlier means that we can only append or prepend columns. We cannot inject them into the middle. I'll give yesod-table a B minus on to composability objective.

Closing Notes and Acknowledgements

One final closing note. You may have noticed that all of the Tables in this post were parameterized over site. This is because they don't depend on a particular foundation type. Usually, the way that this can happen is that you use a route in one of the columns:

-- Assume that our foundation type was named App
peopleTableWithLink :: Table App (Entity Person)
peopleTableWithLink = mempty
  <> peopleEntityFullInfoTable
  <> Table.linked "Profile Page" (const "View") (PersonProfileR . entityKey)

The above example must be parameterized over App (or whatever your foundation type is named), not over site.

This monoidal approach to building tables was inspired by Gabriel Gonzalez's Equational Reasoning at Scale and by the classic diagrams paper. I hope that others find the library useful. I am very open to pull requests and suggestions, so if you have an idea for a convenience function, feel free to open up an issue on the github page.


comments powered by Disqus