A small data structure library of ordered maps and sets in Clojure and ClojureScript.
Add the following dependency to the :deps map in deps.edn:
org.clojars.quoll/tiara {:mvn/version "0.3.7"}
[org.clojars.quoll/tiara "0.3.7"]
At this point, Tiara only includes an ordered map and an ordered set. These have the same O(1) access characteristics as hashmaps/hashsets, but maintain insertion order. However, dissoc
and disj
will be o(n) in the worst case.
Ordered Maps behave the same as other maps, but are only created using the ordered-map
function.
Ordered Sets are the same as other sets, but can be created using ordered-set
when passing the set's contents as arguments, or oset
when passing the contents in a seq.
(require '[tiara.data :refer [ordered-map ordered-set oset]])
;; maps
(def m (ordered-map :a 1 :b 2 :c 3 :d 4 :e 5))
(def h (hash-map :a 1 :b 2 :c 3 :d 4 :e 5))
(= h m) ;; => true
(seq m) ;; => ([:a 1] [:b 2] [:c 3] [:d 4] [:e 5])
(assoc m :f 6) ;; => {:a 1, :b 2, :c 3, :d 4, :e 5, :f 6}
;; but existing keys do not change location:
(assoc m :c 7) ;; => {:a 1, :b 2, :c 7, :d 4, :e 5}
;; to append, then remove first:
(assoc (dissoc m :c) :c 7) ;; => {:a 1, :b 2, :d 4, :e 5, :c 7}
;; sets
(def os (ordered-set :a :b :c :d :e :f :g :h :i))
(def hs (hash-set :a :b :c :d :e :f :g :h :i))
(= os hs) ;; => true
(seq os) ;; => (:a :b :c :d :e :f :g :h :i)
(conj os :j) ;; => #{:a :b :c :d :e :f :g :h :i :j}
(disj os :b) ;; => #{:a :c :d :e :f :g :h :i}
;; adding again does not change location
(conj os :d) ;; => #{:a :b :c :d :e :f :g :h :i}
;; to move to the end, remove and add
(conj (disj os :d) :d) ;; => #{:a :b :c :e :f :g :h :i :d}
I often encounter cases where I need to append data to a structure that will be written out in the same order it was read. However, duplicates of a simple piece of data should not be duplicated, or for more complex data, they should be merged with previously encountered information. These ordered sets and maps are ideal for this operation.
While it would be possible to use Java's LinkedHashMap and LinkedHashSets, these are not immutable structures, and hence are not ideal for functional programming. I suspected that an immutable equivalent may have existed, but the required functionality did not appear difficult and I thought it would be an interesting learning experience to write my own version.
In fact, there were already existing libraries (found by Tom Dalziel):
All 3 implementations use a backing hashmap. Hashmaps use a hash-trie structure to find entries by key, though these are amalgamated over several entries, with one tree node being shared by up to 32 entries. Since these are common to all 3 systems, and incur a small overhead per entry, the overall space used by these maps can be considered as roughly the same between systems.
Tiara uses a backing hashmap and a backing vector. When a key/value is stored, then a MapEntry
is created and stored at the end of the vector. The size of the vector (and hence, the location of this new entry) is then stored with the key in the hashmap.
Each entry therefore uses a standard MapEntry
for the key/location pair in the backing map, and a second MapEntry
for the key/value
pair in the vector.
Looking up a key will access the backing map for the location in the vector, and the key/value pair in the vector can be found at that location. This is a single indirection to the map entry. If the entry is requested via find
then this can be returned directly from the vector.
Removing an entry is an expensive operation. The location in the vector is found, and a new vector is formed from the subvecs before and after that point. However, every index in the hashmap that referred to a location after the removal point needs to be decremented. These entries can be found by the keys from the second subvec, but the operation is still O(n). Removing every entry would be O(n^2). Fortunately, most uses of maps do not require significant removals, so I decided this was acceptable. However, in the case where a lot of removals are necessary, this would not be the right library to use.
Finally, the seq of the map is already available as the vector, so this can be returned with no work required.
This is a similar structure to Tiara, using a hashmap for fast indexing, and a vector to store the order.
The hashmap entries consist of a nested MapEntry
pair. Maps usually store a single key/value entry, but in this case the value is another MapEntry
. This inner MapEntry
contains the value for the key, and the index to find the value in the vector. The vector also contains another MapEntry
object, this time containing the key/value pair. This therefore needs 3 MapEntry
objects per insertion.
Looking up a key will access the backing hashmap to find the inner MapEntry
, and then look up the value inside that object. This is also a single indirection. However, finding a full entry (using find
) requires the construction of a new MapEntry
object to return.
Removing an entry is done by removing the entry from the backing hashmap, and setting the position in the vector to a nil
value. If a lot of removals are done, this will make the vector sparse. To account for this, a new Compactable
protocol was introduced, with a compact
function to remove all the nil
entries and reindex the entries. This allows the cost of removal to be postponed and multiple removals can be compacted at once, though the operation must be manually selected. Tiara could do this, but I have opted to stick with immediate compaction, since most of my use cases do not need this.
The seq of an Ordered map is based on filtering the nils out of the backing vector. The filtering is lazy, and not a significant cost when compared to whatever is using the seq.
This is a significantly different data structure. Each entry is given a node that contains the value, along with references to left
and right
. These are the key values to find nodes before and after them, which is an indirection for creating a doubly-linked list. (Indirection is needed for doubly-linked lists in functional data structures, since second updates to nodes will not update any references to the node). The linked list is circular, to allow iteration both forward and back, starting from the stored head of the list.
Storage requires the default MapEntry
in the backing hashmap, along with a Node
object that will go to the tail position of the list. This node will set the right
pointer to the head, and the left
reference to the previous tail
. The head must set its left
to point to the new tail, and the previous tail must update its right
reference to point to this new tail. Due to indirection, this means looking up the existing head
and tail
and then setting their left
and right
references, respectively. This is a lot of updates to the backing map, and incurs a significant cost.
Looking up a key retrieves the node with the value, which can be returned immediately. If the entry has been requested (with find
) then a new MapEntry
must be constructed.
Removal is similar to insertion, in that the node can be found by a lookup in the tree, and the keys for the nodes to its left
and right
can be found. These nodes can be set to refer to each other, thereby removing the target node from the linked list. Finally, the node can be removed from the backing map.
The ordered seq can be built out of the nodes using their linked list. However, due to each node holding only the key of the next node in the list, this means that every step in the iteration requires a map lookup.
Some simple Criterium tests demonstrate the performance differences between these 3 implementations. To minimize the non-map working being performed, only numbers are being stored as keys and values.
To avoid access patterns, the initial data was generated with:
(def data (shuffle (range 1000000)))
The numbers presented here are on JDK 21, running on a 14-inch MacBook Pro, Nov 2023, with an M3 Max chip. Three benchmarks were executed for each operation on each library. The best time out of the 3 was selected, though the results were generally consistent.
(bench (into (linked.core/map) (map (fn [n] [n n]) data)))
(bench (into (flatland.ordered.map/ordered-map) (map (fn [n] [n n]) data)))
(bench (into (tiara.data/ordered-map) (map (fn [n] [n n]) data)))
Library | Time (ms) | Std Dev (ms) |
---|---|---|
Linked | 839.4 | 36.3 |
Ordered | 365.4 | 106.3 |
Tiara | 330.3 | 33.8 |
Ordered used to be the fastest here, by a slim margin over Tiara. However, Tiara did not include Transient versions of the data Structures. Since these were introduced, the timing has improved by about 30%. Linked was the slowest by a long way, which reflects the multiple updates required for each insertion. Note that Ordered has a standard deviation of 29%.
However, after saving each of these maps (shown in the next test), attempting to run this benchmark on Ordered led to Heap Exhaustion! This problem did not affect either Linked or Tiara. Restarting the repl with a larger heap allowed this test to proceed. This high memory usage also explains the high standard deviation, since heap access (and potentially garbage collection) becomes an issue.
Using the same data, the maps were saved:
(def linked (into (linked.core/map) (map (fn [n] [n n]) data)))
(def ordered (into (flatland.ordered.map/ordered-map) (map (fn [n] [n n]) data)))
(def tiara (into (tiara.data/ordered-map) (map (fn [n] [n n]) data)))
Using the randomized data as keys, each value was looked up and added to an accumulator. This operation therefore accessed every member of the map:
(bench (reduce (fn [acc n] (+ acc (get linked n))) 0 data))
(bench (reduce (fn [acc n] (+ acc (get ordered n))) 0 data))
(bench (reduce (fn [acc n] (+ acc (get tiara n))) 0 data))
Library | Time (ms) | Std Dev (ms) |
---|---|---|
Linked | 119.3 | 5.4 |
Ordered | 113.6 | 4.0 |
Tiara | 118.5 | 5.4 |
All three are relatively close with Tiara being 4% slower than Ordered. This has changed since 2023, when Tiara used to be slightly faster. Ordered used to be quite a bit slower than the other 2, but has since become the fastest.
The reason for this data structure is to be able to return data in order, so the next test was to look at that. Accessing the vals means pulling data out of the ordering, which may be different to getting the full seq, so I looked at both. I started by looking at the vals:
(bench (reduce + 0 (vals linked)))
(bench (reduce + 0 (vals ordered)))
(bench (reduce + 0 (vals tiara)))
Library | Time (ms) | Std Dev (ms) |
---|---|---|
Linked | 332.5 | 15.7 |
Ordered | 48.2 | 1.3 |
Tiara | 18.9 | 0.4 |
Ordered and Tiara are about the same here. The small difference could just be noise, but there did seem to be a consistent difference. This would be due to the wrapping keep identity
operation over the vector that filters out any nil
values in Ordered.
Linked is the outlier, due to its need to access the map in each step along the linked list.
Next, I looked at seqs, iterating over each entry, but ignoring the values:
(bench (reduce (fn [a _] (inc a)) 0 (seq linked)))
(bench (reduce (fn [a _] (inc a)) 0 (seq ordered)))
(bench (reduce (fn [a _] (inc a)) 0 (seq tiara)))
Library | Time (ms) | Std Dev (ms) |
---|---|---|
Linked | 322.2 | 28.5 |
Ordered | 6.0 | 0.1 |
Tiara | 2.4 | 0.1 |
Tiara has an advantage in this test, since it already has the seq ready to return, while Ordered needed to create MapEntry
objects for each element. Ordered still does well though.
When not removing items, Tiara seems to have slightly better performance than Ordered, and quite a lot better than Linked. While slightly slower (4%) in random access, this is not a common use case for this kind of data structure.
Despite no changes in code, previous tests showed Tiara to be better than Ordered for random access. The difference here is the testing platform. While the newer CPU is likely to benefit all platforms equally, the newer JDK shows much better memory management performance. This allows Ordered to run without heap exhaustion (and used to require a JDK flag for an expanded heap before it could run), and has significantly improved the allocation/deallocation of the MapEntry
objects that it uses.
Both Linked and Tiara seem to have better memory consumption profiles than Ordered, though JDK 21 appears to remove this as an issue.
For access patterns that do not require significant removals, then Tiara should provide some benefits.
Copyright © 2023, 2024 Paula Gearon
Distributed under the Eclipse Public License version 2.0.