<< Back to previous view

[CLJ-1388] equality bug on records created with nested calls to map->record Created: 18/Mar/14  Updated: 23/May/14

Status: Open
Project: Clojure
Component/s: None
Affects Version/s: Release 1.6
Fix Version/s: Release 1.7

Type: Defect Priority: Major
Reporter: Nicola Mometto Assignee: Unassigned
Resolution: Unresolved Votes: 0
Labels: defrecord

Attachments: Text File 0001-FIX-CLJ-1388.patch     Text File CLJ-1388.patch     Text File CLJ-1388-record-equality-and-map-record-factory.patch     Text File CLJ-1388v2.patch    
Patch: Code and Test
Approval: Screened

 Description   

Depending on the type of the map passed to a record map constructor, records will not correctly compare for equality:

user> (defrecord a []) 
user.a
user> (def r1 (map->a {:a 1}))
nil
user> (def r2 (map->a r1))
nil
user> (= r1 r2)  ;; expected => true
false
user> (.__extmap r1)
{:a 1}
user> (.__extmap r2)  ;; expected => {:a 1}
#user.a{:a 1}

Cause: The type of the map passed into the map constructor leaks into the __extmap, affecting equality comparison of the record. This bug was described in this post: https://groups.google.com/forum/#!topic/clojure/iN-SPBaTFUw

Approach: Clean the extmap before putting it into the record constructor.

Patch: CLJ-1388-record-equality-and-map-record-factory.patch

Screened by: Alex Miller



 Comments   
Comment by Steve Miner [ 28/Apr/14 3:06 PM ]

The proposed patch 0001-FIX-CLJ-1388.patch makes every map->Record method pay to copy the argument map every time. However, according to my tests, the problem only occurs with records without any fields. So it should be sufficient to generate the (into {} m#) case only when `fields` is empty. [Update: this is wrong, explained below.]

Comment by Steve Miner [ 28/Apr/14 3:10 PM ]

It would be better to fix the problem in the Java Record/create method, but I couldn't figure out how that worked. On the other hand, this bug seems like a fairly rare edge case so I think my patch is acceptable.

Comment by Alex Miller [ 28/Apr/14 3:23 PM ]

Moving out of Screened due to new patch

Comment by Nicola Mometto [ 28/Apr/14 3:35 PM ]

Steve, the problem doesn't occur with records without any fields, your testing was reporting that only because you are only using one record type.

Here's an example that returns true with my patch, but still returns false with yours.

user=> (defrecord a [a])
user.a
user=> (defrecord b [b])
user.b
user=> (def x1 (map->a {:a 1 :b 2}))
#'user/x1
user=> (def x2 (map->a (map->b {:a 1 :b 2})))
#'user/x2
user=> x1
#user.a{:a 1, :b 2}
user=> x2
#user.a{:a 1, :b 2}
user=> (= x1 x2)
false
user=> (.__extmap x2)
#user.b{:b 2}
Comment by Nicola Mometto [ 28/Apr/14 3:37 PM ]

It should also be noted that the overhead of copying the record map is probably insignificant.

Comment by Nicola Mometto [ 28/Apr/14 3:42 PM ]

I also thought at first to fix the problem either on the /create method or on the 3-arity ctor but given that:

  • a fix there would involve messing with the bytecode emitted and thus would be harder to implement than this simple 1-line patch
  • neither the /create method nor the 3-arity ctor is documented and thus should be considered implementation details

I think patching the map->record function is the best way to go.

Comment by Steve Miner [ 28/Apr/14 3:56 PM ]

Nicola, thanks for the correction. I missed the case with multiple records. I withdrew my patch. I'd still like to find a more finely tuned patch, but first I'll have to improve my tests as you demonstrated.

Comment by Steve Miner [ 29/Apr/14 10:17 AM ]

Attached CLJ-1388-record-equality-and-map-record-factory.patch that checks arg to map->Record for MapEquivalence, uses (into {} m#) when necessary. This makes equiv test work correctly with records as the argument (and other map-like values). Added tests with variety of args to map->Record.

Comment by Steve Miner [ 29/Apr/14 10:46 AM ]

A few comments about the new patch... I think the basic issue is a bad interaction between = for records and the generated Record/create method. Everything works when the interal __extmap is a regular map (MapEquivalence), but it fails if __extmap is another record. I think that's because of equiv calling = on the __extmap's.

The user expects to create a new record using the value of another record because it's just like a map. However, = on records respects the record type so it's not = to a map.

The general work-around is to use (into {} x) on the argument to the map->Record. To meet the user's expectation, that `into` call can be incorporated into the map->Record. But I didn't like the defensive copy as most of the time it's unnecessary – the argument is typically a regular map. The `into` work-around is only necessary if the arg is not a MapEquivalence.

There might be a better way to fix the Record/create method but I couldn't figure it out.

Comment by Nicola Mometto [ 29/Apr/14 1:52 PM ]

Steve's last comment made me realize that the root of the problem is on the record .equiv method, where the extmaps are compared via `=`

This new patch (CLJ-1388.patch) addresses this issue by comparing the extmaps with Utils/equiv rather than `=`, which compares maps in a type-indipendent way.

There's still a case where we need recreate the given map, that is when the given map is not an IPersistentMap but simply a java.util.Map.

Steve, my new patch incorporates my fix and your tests, I modified your patch to include only the tests (that were really comprehensive) since I figured it's fair to keep your authorship on those, let me know if that's a problem with you.

Comment by Steve Miner [ 29/Apr/14 2:10 PM ]

Whatever works for you regarding the tests is fine by me.

Comment by Alex Miller [ 30/Apr/14 12:07 AM ]

It seems weird to me that a record should ever contain another record as its extmap. We should be considering the performance aspect but I'm concerned that not locking down extmap more just invites other weirder problems later.

In CLJ-1388.patch, you mention Utils/equiv in your comment but the patch calls Utils/equals - which did you mean?

Also, that patch currently checks if m# is an IPersistentMap - I can't imagine what case we would want to allow where a valid m# is NOT an IPersistentMap?

Comment by Nicola Mometto [ 30/Apr/14 4:15 AM ]

Alex, the Utils/equiv in my comment is wrong (it's easy to confuse between equiv/equals, sorry), Utils/equals in the patch is the right method to use since it compares in a type agnostic way.

Since __extmap is an implementation detail and is only used internally by defrecord for its methods, I don't think it's going to be a problem whether it's a record or a regular clojure map. (Clojure only requires it to be an IPersistentMap)

Regarding the check for m# being an IPersistentMap, Steve in his tests had a case where the map->record ctor was invoked with a java.util.Map, I went to look into the docs for defrecord and it only mentions that the argument to map->record has to be a "map", it doesn't specify that it has to be a clojure map/IPersistentMap, so it seemed right to allow for java maps too and wrap them in IPersistentMaps internally.

Comment by Steve Miner [ 30/Apr/14 8:27 AM ]

My test with java.util.Map was an extension of the idea that anything map-like could be used to initialize a record. That might be a bridge too far, but my patch was testing for MapEquivalence to handle records so it made sense to allow j.u.Map, etc. With Nicola's latest patch, it's probably unnecessary to support non-IPersistentMaps so map->Record doesn't actually need to change.

Comment by Nicola Mometto [ 30/Apr/14 3:57 PM ]

CLJ-1388v2.patch is like CLJ-1388.patch except it doesn't copy non IPersistentMaps in a clojure map.

To summarize, here's the status of the different patches for this ticket:

  • 0001-FIX-CLJ-1388.patch copies the argument of map->record in a clojure map via `(into {} m#)`, be it already a clojure map, a record, or a java.util.Map
  • CLJ-1388-record-equality-and-map-record-factory.patch adopts the same approach except it only copies the arg of map->record into a clojure map if the arg doesn't satisfy clojure.lang.MapEquivalence
  • CLJ-1388.patch fixes the issue by changing the function that compares __extmaps from `=` (type aware) to `clojure.lang.Utils/equals` (type agnostic), this patch also copies the arg of map->record into a clojure map if the arg doesn't satisfy clojure.lang.IPersistentMap
  • CLJ-1388v2.patch is the same as CLJ-1388.patch except it doesn't copy the arg of map->record into a clojure map if the arg doesn't satisfy clojure.lang.IPersitentMap, thus map->record will not work with bare java.util.Maps (which is the behaviour it has already)
Comment by Alex Miller [ 05/May/14 1:59 PM ]

Are these patches all still in play? Having 4 active patches does not help move a ticket forward.

Can someone re-summarize at this point what questions exist?

Comment by Nicola Mometto [ 06/May/14 5:26 AM ]

0001-FIX-CLJ-1388.patch should be superseded by the other 3 patches since they solve the same problem in a more performant way.

To pick between the other patches, we need to chose which approach to go with.
Patches CLJ-1388.patch and CLJ-1388v2.patch fix the issue in the equiv method of the defrecord, patch CLJ-1388-record-equality-and-map-record-factory.patch fixes the issue in the map->record ctor by converting maps that don't implement MapEquivalence to a clojure map.

I'd go with either CLJ-1388.patch or CLJ-1388v2.patch since they both avoid copying alltoghether in the cases where map->record currently works, while CLJ-1388-record-equality-and-map-record-factory.patch needs to copy the arg into a map if the arg is a custom IPersistentMap or a record.

To pick between CLJ-1388.patch or CLJ-1388v2.patch we need to decide whether or not the current behaviour of map->record to require strictly an IPersistentMap is the way to go: if we decide that it's ok to pass non IPersitentMap maps like java.util.Map to map->record then pick CLJ-1388.patch otherwise CLJ-1388v2.patch

Comment by Alex Miller [ 06/May/14 10:22 AM ]

From brief conversation with Rich, we should not allow arbitrary map types in __extmap so would prefer to force a clean map and rely on standard equality checking. I think CLJ-1388-record-equality-and-map-record-factory.patch is the preferred path based on that, which still seems like it should avoid copying in nearly all common cases.

Comment by Alex Miller [ 23/May/14 11:19 AM ]

Screened specifically CLJ-1388-record-equality-and-map-record-factory.patch - use map as is if it supports MapEquivalence (and can thus be compared under a map) and otherwise dump into a clojure map.

Generated at Thu Jul 31 00:22:38 CDT 2014 using JIRA 4.4#649-r158309.