Vinyl is a general solution to the records problem in Haskell using type level strings and other modern GHC features, featuring static structural typing (with a subtyping relation), and automatic row-polymorphic lenses. All this is possible without Template Haskell.

First, install Vinyl from Hackage:

cabal update
cabal install vinyl

To learn more, try this tutorial.

Vinyl is an approach to records in Haskell based on a heterogeneous list type: a list where each element may have a different type. Since it is a list, accessing any individual field of a Vinyl record seems as though it should have O(n) complexity, with n the size of the record. Sure enough, pattern matching on a Vinyl record works just as with a list: we peel off the head element until we reach the one we are looking for. But a Vinyl record is not a regular list: its type encodes both its length and the type of each of its elements. This means that if we can serialize each field of the record, using something like the Storable type class, then we can serialize the entire record. A Storable instance requires that we statically know the storage size in bytes of values of the type in question, which means that for a collection of Storable values as you might find in a Vinyl record, we know at compile time the byte offset of a particular field given a Ptr to the record as a whole. Finally, then, we can consider a benchmark of getting specific fields from different varieties of Vinyl records.

This benchmark suggests the trends we should expect

• Getting a field deeper in a Vinyl record takes more time than getting one earlier (i.e. it appears to the left, or under fewer :& constructors) in the record
• Different varieties of Vinyl records perform similarly (here we see FieldRec that has Symbol labels of each field, and HList whose field types are unadorned)
• Getting the field of a Vinyl record by composing the field accessor with a Storable peek is both very fast and has performance independent of the field's position in the record

The x-axis of this graph should be labeled as nanoseconds, but the Storable benchmarks all complete in about 450 picoseconds.