Invented by Mapbox, they are a combination of the ideas of finite-sized tiles and vector geometries. Mapbox maintains the official implementation spec for VectorTile codecs.
VectorTiles are advantageous over raster tiles in that:
- They are typically smaller to store
- They can be easily transformed (rotated, etc.) in real time
- They allow for continuous (as opposed to step-wise) zoom in Slippy Maps.
Raw VectorTile data is stored in the protobuf format. Any codec implementing the spec must decode and encode data according to this .proto schema.
vectortiles
is a minimum viable implementation of Version 2.1 of the
VectorTile spec. It aims to be a solid reference from which to implement
other codecs. It exposes a small API of conversion functions between raw
protobuf data and a higher-level VectorTile
type that is more condusive to
further processing. It also exposes fairly simplistic (yet sensible)
implementations of the typical GIS Geometry
types:
- Point
- LineString
- Polygon
For ease of encoding and decoding, each Geometry
type and its Multi
counterpart (i.e. Multipoint) are considered the same thing, a Vector
of
that Geometry
.
This library is not micro-optimized, but does leverage some "for-free" aspects of Haskell to remain usable:
Point
is implemented as a Record Pattern Synonym to hide the fact it's just a vanilla tuple ofInt
s. This allows us to use the more efficient unboxedVector
s with it:
-- | Access a Point's values with the `x` and `y` functions.
type Point = (Int,Int)
pattern Point :: Int -> Int -> (Int, Int)
pattern Point{x, y} = (x, y)
- Some types (like
LineString
) are implemented as anewtype
for its compile-time unboxing:
import qualified Data.Vector.Unboxed as U
newtype LineString = LineString { lsPoints :: U.Vector Point }
- All lenses are
INLINE
d.
You can run the benchmarks with stack bench
, provided you have the stack
tool. The following results
are from a 2016 Lenovo ThinkPad Carbon X1 with an Intel Core i7 processor,
comparing this library with a Python library of similar
functionality. All
benchmarking code is available in the bench
directory.
Note: 1 ms = 1000 μs
One Point | One LineString | One Polygon | roads.mvt (40kb, 15 layers) | |
---|---|---|---|---|
CPython 3.5.2 | 63 μs | 70 μs | 84 μs | 76 ms |
PyPy 5.3 | 116 μs | 210 μs | 211 μs | 12 ms |
Haskell | 3.6 μs | 5 μs | 5.8 μs | 17.1 ms |
The Haskell times are measuring data evaluation to their Normal Form (fully evaluated form).
The Python class decoded to is the builtin dict
class.
One Point | One LineString | One Polygon | roads.mvt | |
---|---|---|---|---|
CPython 3.5.2 | 218 μs | 278 μs | 703 μs | N/A |
Haskell | 3.2 μs | 4.4 μs | 5 μs | 11.1 ms |
Certain encoding benchmarks for Python were not possible.
One Polygon | roads.mvt (water layer) |
|
---|---|---|
CPython 3.5.2 | 84 μs | 78 ms |
PyPy 5.3 | 31 μs | 7.9 ms |
Haskell | 3.4 μs | 6.8 ms |
The operation being benchmarked is ByteString -> Polygon
, meaning we
include the decoding time to account for speed gains afforded by laziness.
- Laziness pays off. In Haskell, just fetching some specific data field is faster than decoding the entire structure.
- Python data fetches are fast. They are based on the
dict
class, so fetch operations will be as fast asdict
is. - PyPy results are enigmatic. Python3 seems to do much better "off the block", but given time the PyPy JIT overtakes it. Fetching layer names also seems to be faster than decoding the entire object, somehow. This may be due to the JIT being clever, noticing we aren't using the rest of the structure.
Simply parsing raw protobuf data is not enough to work with VectorTiles, since the spec also defines how said data is to be interpreted once parsed. In writing a codec, there are a number of things one must consider:
Many languages have a "protobuf compiler" which can take a .proto
file and
generate schema code to access parsed data. There are PROs and CONs to taking
this approach.
PROs for using a protoc
-like program:
- All accessor code is written for you
- Update process when new official
.proto
is released is clearer
PROs for writing your own schema:
- Its your code, so you have more control. Bugs are easier to chase
- The code will likely be much shorter
In the case of two Haskell protobuf libraries which were compared, the hand-written one allowed for a 50-line schema, while the other auto-generated a 550-line one.
The protobuf spec leaves room for additional:
Value
types in the key-value metadata maps- fields in a
Layer
- fields in a
Tile
- use of the
UNKNOWN
geometry type
In writing a codec, you are completely free to ignore these. However, they are permitted by the spec, and some tools may encode data using them.
At the protobuf level, Features of Points, LineStrings, and Polygons are all mixed
into a single list, distinguished only by a GeomType
label. At a high level, you may
wish to separate these specifically. This library does just that:
data Layer = Layer { _version :: Int
, _name :: Text
, _points :: V.Vector (Feature Point)
, _linestrings :: V.Vector (Feature LineString)
, _polygons :: V.Vector (Feature Polygon)
, _extent :: Int
}
As opposed to having a single field named features
, which contains all
features unified by some superclass / trait / generic.
Claim: Having separate accessors for each geometry type yields a "heavier" API, but gives more power, is more performant, and less complex.
Layers and Features have coupled data at the protobuf level. In order to achieve higher compression ratios, Layers contain all metadata in key/value lists to be shared across their Features, while those Features store only indices into those lists. As a result, functions converting protobuf-level Feature objects into a high-level type need to be passed those key/value lists from the parent Layer. and a more isomorphic:
feature :: Geometry g => RawFeature -> Either Text (Feature g)
is not possible.
Version 2 of the spec mainly clarified language surrounding how polygons should be decoded. This Github issue reports another "gotcha" associated with the definition of polygons.
The protobuf data can be malformed in a number of ways. How much sanity checking you wish to do while decoding depends on how much performance you can sacrifice. For instance, here is a constraint found in the spec regarding feature metadata:
Every key index MUST be unique within that feature such that no other attribute pair within that feature has the same key index. A feature MUST have an even number of tag fields. A feature tag field MUST NOT contain a key index or value index greater than or equal to the number of elements in the layer's keys or values set, respectively.
Decode what you encode, and encode what you decode.
Your encoding and decoding functions should be as close to isomorphisms as possible.
(0,0) is in the top-left corner.
Know your binary arithmetic.
Lists of Geometry commands/values are Z-encoded. See the zig
and unzig
functions in Geometry.VectorTile.Geometry
.