I'm a bit late to this, but it caught my eye.
The only common use case I have for mutating data structures in Clojure is
when storing state in a global map (similar to Om.Next), but I almost
always make them atomic to account for nondeterminism in the order
operations on them will finish. Would the performance gains of Bifurcan's
Map and IntMap over Clojure's PersistentHashMap hold for atomic versions?
Faster set operations might be useful for me as well, since I often end up
rolling my own, although less so for maps and more for vectors with a small
number of values so not sure whether that's applicable here.
Thanks,
Sophia
On Friday, April 21, 2017 at 12:53:56 AM UTC-4, Zach Tellman wrote:
>
> Sure, happy to elaborate. Bifurcan offers potential performance wins a
> few different ways:
>
> * We can use standard Java equality semantics, bypassing all the overhead
> of the hash calculations and enhanced numeric equality checks (this can
> lead to moderate performance gains)
> * We can use a mutable data structure as long as it never escapes a local
> context (this can lead to significant performance gains)
> * We can use the extra capabilities the data structures expose, like
> concatenation, slicing, set operations, etc. (this is too dependent on the
> use case to really quantify)
>
> it would be easy to have a `map` and `map*` method that expose Clojure and
> Java equality semantics, respectively, but that puts a big onus on the
> developer to determine if the latter is safe for their use case. I've been
> bit by this when I've used j.u.c.ConcurrentHashMap before, so I expect
> people will suffer similarly weird bugs.
>
> However, I think there's a way to use the mutable data structures.
> Technically, transient data structures allow arbitrary persistent data
> structures to be batch updated, but in practice they tend to be empty, and
> after they're populated they tend to be treated as read-only.
>
> If we're convinced this is common enough, every empty transient data
> structure could be mutable, and when we make it persistent we could wrap it
> in a "virtual" collection [1] which allows updates without touching the
> base collection. This would allow for faster writes, faster reads, and
> only marginally slower updates if those are required.
>
> This is all predicated on a bunch of assumptions that are hard to
> validate, but if this describes enough real-world use cases, it could lead
> to a big, easy performance win. It's even possible to automatically
> replace the base Clojure collections with these alternatives using
> something like Sleight [2].
>
> Anyway, that's what I've been mulling over. If anyone has opinions, I'm
> happy to hear them.
>
> Zach
>
> [1]
> https://github.com/lacuna/bifurcan/blob/master/src/io/lacuna/bifurcan/Maps.java#L103
> [2] https://github.com/ztellman/sleight
>
> On Thu, Apr 20, 2017 at 8:55 AM Dave Dixon > wrote:
>
>> Sounds great. If you have time, I'd certainly like to hear your thoughts
>> on the issues of equality semantics and transients, maybe I can ponder and
>> make some suggestions based on my target use-case.
>>
>>
>> On Tuesday, April 18, 2017 at 9:32:32 AM UTC-7, Zach Tellman wrote:
>>
>>> To be clear, my intention was always to wrap the implementations in the
>>> appropriate Clojure interfaces, and I don't believe that will cause much,
>>> if any, of a performance hit (inlining is magic). However, there are some
>>> real questions regarding how to expose non-standard equality semantics, and
>>> whether transients should be represented using the immutable or mutable
>>> collection variants.
>>>
>>> For what it's worth, I have about 1/3 of an implementation of
>>> Clojure-compatible versions of these data structures, I just wanted to mull
>>> on the above questions a bit before going further. I'm happy to discuss
>>> them here in more depth if you have any questions or opinions.
>>>
>>> Zach
>>>
>>> On Tue, Apr 18, 2017 at 6:53 AM Dave Dixon wrote:
>>>
>> Stared at this a bit yesterday. Seems like if you want to leverage spec
while using bifurcan, then the bifurcan types need to have the Clojure
wrapper. The alternative appears to be re-implementing at least a large
subset of collection-related spec code, which is a lot to bite off. Also
tried updating some existing code to use bifurcan. Similar to spec, there
are going to be cases which are less perf sensitive, where it would be
nice
to use code that is polymorphic for collections, and drop down to the fast
interface in perf-sensitive parts.
On Monday, April 17, 2017 at 1:52:39 PM UTC-7, Dave Dixon wrote:
>
> What is the issue with wrapping in Clojure interfaces? Added overhead
> of function calls?
>
> I'm finding myself in the process of doing some of this, at least for
> constructors. Also thinking of generating predicates/generators for use
> with spec.
>>>
