Skip to content

Latest commit

 

History

History
237 lines (179 loc) · 8.38 KB

README.md

File metadata and controls

237 lines (179 loc) · 8.38 KB

DataFrame and Numerics for ClojureScript

Clojars Project

Minimal cljs implementation of tech.v3.datatype, tech.v3.datatype.functional, tech.v3.datatype.argops, and tech.v3.dataset. This implementation is based on typed-arrays for numeric data and js arrays for everything else so it should support all your favorite datatypes. Support for columns of type java.time.Instant and java.time.LocalDate is included.

Datasets serialize and deserialize much faster than sequences of maps. They use less memory and they allow faster columnwise operations. To transform a sequence of maps into a dataset use ds/->dataset. To get a sequence of maps back use ds/rows.

This library is mainly going to useful if you are dealing with large amounts of primarily numeric data such as timeseries data coming off of a sensor. In that case you can specify exactly the datatype of the column which will get you major benefits in terms of memory and serialization size. I developed this library when working with such data in a react-native application.

Unlike the jvm-version this is a very minimal exposition of these concepts. Since the underlying vm itself is typeless there was no need for a complex macro system to do unboxed math in loops so I could stay much closer to core clojure and in fact ICounted and IIndexed are the primary interfaces and tech.v3.datatype/reify-reader creates a persistent-vector hash and equiv compatible object.

If you are totally new to all of this please check out my talk on High Performance Data with Clojure. Many more talks are available from my GitHub page.

Example

Server Side

There is a new namespace, tech.v3.libs.transit that contains a transit-safe dataset->data function. There are also transit handlers defined if you know how to override your transit handlers in your middleware.

(defn generate-data
  [request]
  (-> (ds/->dataset (repeatedly 10000 #(hash-map :time (rand)
                                                 :temp (rand)
                                                 :temp1 (rand)
                                                 :temp2 (rand)
                                                 :valid? (if (> (rand) 0.5)
                                                           true
                                                           false))))
      (ds-t/dataset->data)
      (response/response)))

Client Side

The same namespace tech.v3.dataset is exposed for clojurescript containing most of the usual functions - columns, rows, select-rows, group-by, etc. The version of dataset->data in this namespace corresponds with the version on the jvm side above. There are also transit handlers defined if you have access to override the transit handlers in your middleware stack.

  (GET "/data" {:handler #(let [ds (ds/data->dataset %)]
                            (swap! app* assoc
                                   :ds ds
                                   :raw (mapv (partial into {}) (ds/rows ds))))})

Quick Walkthough

;;Create a dataset from a map of columns
cljs.user> (def ds (ds/->dataset {:a (range 100)
                                  :b (take 100 (cycle [:a :b :c]))
                                  :c (take 100 (cycle ["one" "two" "three"]))}))

#'cljs.user/ds
cljs.user> ds
#dataset[unnamed [100 3]
| :a | :b |    :c |
|---:|----|-------|
|  0 | :a |   one |
|  1 | :b |   two |
|  2 | :c | three |
|  3 | :a |   one |
|  4 | :b |   two |
...

;; Control column datatypes by using parser-fn which is very thoroughly documented in
;; tech.ml.dataset api documentation

cljs.user> (->> (ds/->dataset {:a (range 100)
                               :b (take 100 (cycle [:a :b :c]))
                               :c (take 100 (cycle ["one" "two" "three"]))}
                              {:parser-fn {:a :int8}})
                (vals)
                (map (comp :datatype meta)))
(:int8 :keyword :string)


;; Datasets are always safe to print to your repl.  Only the first 25 rows are printed.
;; You can use datasets like maps of columns - columns also are safe to print to your
;; repl at all times.
cljs.user> (ds :a)
#column[[:float64 100][0 1 2 3 4 5 6 7 8 9 ... 90 91 92 93 94 95 96 97 98 99]

;;Add a new column with map.
cljs.user> (ds/head (assoc ds :aa (map #(* % %) (ds :a))))
#dataset[unnamed [5 4]
| :a | :b |    :c | :aa |
|---:|----|-------|----:|
|  0 | :a |   one |   0 |
|  1 | :b |   two |   1 |
|  2 | :c | three |   4 |
|  3 | :a |   one |   9 |
|  4 | :b |   two |  16 |]


;; Column map is bit more efficient
cljs.user> (ds/head (ds/column-map ds :aa #(* % %) [:a]))
#dataset[unnamed [5 4]
| :a | :b |    :c | :aa |
|---:|----|-------|----:|
|  0 | :a |   one |   0 |
|  1 | :b |   two |   1 |
|  2 | :c | three |   4 |
|  3 | :a |   one |   9 |
|  4 | :b |   two |  16 |]

;; Remove columns with dissoc
cljs.user> (ds/head (dissoc ds :c))
#dataset[unnamed [5 2]
| :a | :b |
|---:|----|
|  0 | :a |
|  1 | :b |
|  2 | :c |
|  3 | :a |
|  4 | :b |]

;;select particular rows
cljs.user> (ds/select-rows ds [1 3 5 7 9])
#dataset[unnamed [5 3]
| :a | :b |    :c |
|---:|----|-------|
|  1 | :b |   two |
|  3 | :a |   one |
|  5 | :c | three |
|  7 | :b |   two |
|  9 | :a |   one |]
  • Once you have a dataset using the function in the dataset namespace to do columnwise operations will usually be a bit quicker.
  • See testapp for a minimal quick walkthrough and verification that advanced optimizations do not break the api.

Chrome Heap Measurements.

For a 2-column dataset of just time and temperature of random double data dataset is about 18X smaller than a sequence of maps. More columns or choosing different datatypes for the columns will change the results in various ways but 10X isn't unreasonable.

testapp.webapp> (def ignored
                  (->> (repeatedly 10000 #(hash-map :time (rand) :temp (rand)))
                       (ds/->>dataset)
                       ;;When building a dataset we save data using dynamically resizing
                       ;;backing stores.  Cloning crops the dataset to exactly the size
                       ;;it needs
                       (clone)
                       (aset js/window "AAAMemTest-Dataset")))
#'testapp.webapp/ignored
testapp.webapp> (def ignored (->> (repeatedly 10000 #(hash-map :time (rand) :temp (rand)))
                                  (vec)
                                  (aset js/window "AAAMemTest-Mapseq")))
#'testapp.webapp/ignored

heap profile

Gotchas

The js :uint64 and :int64 typed arrays produce BigNum objects. When setting values I transparently upcast to bignum but when getting values back you get bignums so be careful. I suggest avoiding using them unless you are sure of what you are doing. You can cast a bignum back to a number, with potential numeric loss, with (js/Number. bn). This is unfortunate because the default clojure number type on the jvm is a long and a dataset of longs has a column of type :int64 and thus if you naively send it over the wire you will have issues trying to work with that dataset in your code.

Development

This is what I have so far to make development quick

Get a REPL

  • clj -M:cljs node-repl
  • cider-connect to port 8777 once it starts
  • (shadow/repl :node-repl)

Unit Tests

There is a test script - scripts/run-tests that does:

  • clj -M:cljs compile test
  • node target/test.js

Install locally and try on different project

  • scripts/install-local

Talk About It

License

  • MIT