[erlang-questions] Maps

Joe Armstrong erlang@REDACTED
Mon May 13 12:15:03 CEST 2013

On Thu, May 9, 2013 at 10:56 PM, Robert Virding <
robert.virding@REDACTED> wrote:

> Even before taking a really deep dive into studying the EEP one thing I
> can immediately say: get rid of this having both *equal* and *match* and *USE
> ONLY MATCH. *Keys are the same when they match. Period. This paragraph:
> If key K does not *equal* any existing key in the map, a new association
> will be created from key K to value V. If key K is *equal* to an existing
> key in map M its associated value will be replaced by the new value V *and
> * the new key K will replace the old key if the key does not *match*. In
> both cases the evaluated map expression will return a new map.
> is weird. Yes I know what it means but it is not intuitive. When keys are
> replaced or not replaced when they are equal is not can seem very strange
> to those not deep into erlang semantics.

I think the problem here is the description - not the semantics. It's not
the keys which are replaced, the crucial
idea is to have two different syntaxes. This needs a longer explanation to
make sense.

Short explanation

    ':=' means "update an existing key - crash if they key is not present"
    '=>' means "update an existing key OR add a new key"

The value is pretty much irrelevant

Long explanation of why this is good

We can update an existing map M with the syntax:

    M#{ K1 Op V1, K2 Op V2, ... Kn Op Vn}

Where Op is either => or :=.

The syntax K => V never generates an error and is used to introduce a new
or to update an existing key with a new value.

The syntax K :=  V is used to update the value of an *existing* key and will
raise an exception if the key K does not already exist in the map M.

The difference between these two modes of update is crucial, but needs
a couple of examples to explain:

Assume we define a map M as follows:

    M = #{ foo => 1, bar => 2, baz => 3 }

The update

    M # {foo := 12,  bary := 24}

Will fail (raise an exception) since there is no key called bary in the
original map. This is good idea (ROK suggested this in his frames paper)
since we don't want too accidentally create a new key due to a spelling
error. This is the crash-early property of the := update syntax.

Crashing later would make debugging difficult, we would accidentally add
a bad key to a map and learn about it way later.

The update

   M # {foo := 12, bar := 24}

Will succeed, but more importantly the new map has exactly the same
keys as the old map (since all the updates are ':=' updates) - and so
can *share* the same key descriptor. So if we have a very long list of
maps they can be stored in a space efficient manner. (Again this idea
comes from ROKs frames paper). Björn-Egil's eep didn't mention this
but the fact that we know that two maps have the same keys from the
syntax make a lot of optimizations possible).

All of this is possible because there are two operators not one :-)


As regards efficiency, utility, beauty and so on these are subjective.

If you want the last ounce of efficiency records and dicts are not
going to go away when maps arrive. So if maps have the wrong
performance characteristics then use the exiting mechanisms.

In the latest addition of my book I've been documenting the changes to maps
- this chapter has changed three times and has tended to be conservative
so I haven't (yet) mentioned that keys can be any term (and not just atoms).

I'm rather looking forward to being able to represent things like XML
and JSON and property lists in a maps and to have an one-size-fits-all
replacement for dicts and records. I've never really worried about the
last ounce of efficiency - if I want real efficiency I'd change
language and go to C or program an FPGA.



> Push
> Yes, I think we made an error with the different types of equalities and
> that comparisons covert integers to floats. With 20-20 hindsight I would
> prefer ==,/=,<,=<,>,>= to *ONLY* work on numbers and convert while
> another set @==,@/=,@<,@=<,@>,@>= (say) to work on terms and never convert.
> Pop
> While we are at it I think the evaluation order for keys and values should
> be defined as left-to-right. The rest of erlang is so why not here?
> Robert
> ------------------------------
> *From: *"Björn-Egil Dahlberg" <egil@REDACTED>
> Hi everyone!
> We finally have a Maps EEP for you. This will get you an idea on what we
> are working on. Some of the ideas presented here was presented at Erlang
> Factory SF Bay Area 2013.
> I am sure that there will be some discussions about the contents of this
> EEP and I hope the discussions will be both fruitful and civilized.
> The journey of Maps and this EEP has been long and by no means a
> straight-forward or continuous one. I had a crystal clear picture of what I
> wanted Maps to be when we first started discussing it within OTP about
> two-three years ago. This EEP resembles that vision but it has had a lot of
> contributions of other ideas from both within and outside of OTP.
> The idea was a functional data-type, a syntax aware mapping of key-value
> associations with pattern matching. A syntax similar to records but without
> the hazzle of compile-time dependency and with arbitrary terms as keys.
> Order was not important and it could be implemented with a
> Hash-Array-Mapped-Trie with good performance and memory trade-offs. This
> was not an approach to replace records. It was meant to replace records
> where suitable and in other regards not be a replacement but its own **
> thing**.
> From the community there has been many wishes of a Map like data-type and
> a few suggestions.  The one suggestion that stands out is of course the
> Frames proposal from Richard O'Keefe. It is the most complete proposal I've
> seen and is very well thought out. Its goal is to be a record replacement
> and the proposal satisfies this goal very well.
> - If Frames are that good, why a separate EEP?
> - It boils down to goals and constraints.
> A record replacement is just that, a replacement.
> It's like asking the question, "What do we have?" instead of "What can we
> get?"
> The instant rebuttal would be "What do we need?" I say Maps.
> Frames has certainly influenced Maps. In many regards Maps also
> encompasses Frames but Maps tries to do more. I think the most significant
> difference would be, arbitrary terms as keys and how many different keys we
> would have in a Map. In the end I believe they are two different things and
> have different goals.
> Some Notes and Disclaimers:
> Later iterations of Maps has gone through some changes, most
> significantly,
>   * From a set of key-values to a ordered set of key-value associations
> I was originally against this change since it forces restrictions on the
> implementation and it illuminates the lack of distinction between
> arithmetic order and term order, i.e. the problem of mixing integer and
> float types as keys in a tree. However, I was later persuaded that key
> ordering is necessary. We have to respect the totalitarian order of terms.
> Considerations has been made on how to, if at all possible, apply Frames
> characteristics to Maps. Most significantly memory space and key-sharing
> characteristics. This is not detailed in the EEP though, just mentioned.
> The function interface has had many revisions as well. At some stage the
> API was considered to be a drop-in-replacement for `dict` and thus would
> have the same function-names. This goal/constraint was dropped by Technical
> Board decision recently.
> From the very beginning Maps was envisioned to have the ability to bind
> variables derived from the environment. Like this:
>     function(K1, #{ K1 := K2, K2 := V }) -> V.
> This feature is a beast. Powerful and scary. It is not confined to only
> Maps but should also be applicable to other types as well:
>     function(Skip, <<_:Skip/binary, Value:Size, _/bits>>, Size) -> Value.
> It is uncertain how effective such an implementation would be and in the
> end we might not want this feature at all.
> In this EEP we will describe syntax and semantics of Maps but very little
> is disclosed of its actual implementation. Current prototypes stems from
> using sparse tuples in a HAMT-like data structure and tuple-like data
> structures. The HAMT-like data structure is discontinued and will be
> replaced by some ordered tree.
> The proposal is included as an attachment but can also be viewed at this
> git-repository:
> https://github.com/psyeugenic/eep/blob/egil/maps/eeps/eep-0043.md
> Regards,
> Björn-Egil Dahlberg
> Erlang/OTP
> _______________________________________________
> erlang-questions mailing list
> erlang-questions@REDACTED
> http://erlang.org/mailman/listinfo/erlang-questions
> _______________________________________________
> erlang-questions mailing list
> erlang-questions@REDACTED
> http://erlang.org/mailman/listinfo/erlang-questions
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20130513/d5c33b61/attachment.htm>

More information about the erlang-questions mailing list