<< Back to previous view

[CLJ-1239] faster, more flexible dispatch for clojure.walk Created: 29/Jul/13  Updated: 26/Jan/16

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

Type: Enhancement Priority: Critical
Reporter: Stuart Sierra Assignee: Stuart Sierra
Resolution: Unresolved Votes: 8
Labels: walk

Attachments: Text File 0001-CLJ-1239-protocol-dispatch-for-clojure.walk.patch     Text File 0002-CLJ-1239-protocol-dispatch-for-clojure.walk.patch    
Patch: Code
Approval: Triaged


The conditional dispatch in clojure.walk is slow and not open to extension. Prior to CLJ-1105 it did not support records.

Approach: Reimplement clojure.walk using protocols. The public API does not change. Users can extend the walk protocol to other types (for example, Java collections) if desired. As in CLJ-1105, this version of clojure.walk supports records.

Patch: 0002-CLJ-1239-protocol-dispatch-for-clojure.walk.patch

Performance: My tests indicate this is 1.5x-2x the speed of the original clojure.walk. See https://github.com/stuartsierra/clojure.walk2 for benchmarks.

Risks: This approach carries some risk of breaking user code that relied on type-specific behavior of the old clojure.walk. When running the full Clojure test suite, I discovered (and fixed) some breakages that did not show up in clojure.walk's unit tests. See, for example, commit 730eb75 in clojure.walk2

Comment by Vjeran Marcinko [ 19/Oct/13 12:32 PM ]

It looks, as it is now, that walking the tree and replacing forms doesn't preserve original meta-data contained in data structures.

Comment by Andy Fingerhut [ 23/Nov/13 1:11 AM ]

Patch 0001-CLJ-1239-protocol-dispatch-for-clojure.walk.patch no longer applies cleanly to latest Clojure master since the patch for CLJ-1105 was committed on Nov 22, 2013. From the description, it appears the intent was either that patch or this one, not both, so I am not sure what should happen with this patch, or even this ticket.

Comment by Alex Miller [ 23/Nov/13 1:52 AM ]

This patch and ticket are still candidates for future release.

Comment by Stuart Sierra [ 20/Dec/13 9:14 AM ]

Added new patch that applies on latest master after CLJ-1105.

Comment by Chouser [ 27/Feb/14 10:26 AM ]

The way this patch behaves can be surprising compared to regular maps:

(clojure.walk/prewalk-replace {[:a 1] nil} {:a 1, :b 2})
;=> {:b 2}

(defrecord Foo [a b])
(clojure.walk/prewalk-replace {[:a 1] nil} (map->Foo {:a 1, :b 2}))
;=> #user.Foo{:a 1, :b 2}

Note how the [:a 1] entry is removed from the map, but not from the record.

Here's an implementation that doesn't suffer from that problem, though it does scary class name munging instead: https://github.com/LonoCloud/synthread/blob/a315f861e04fd33ba5398adf6b5e75579d18ce4c/src/lonocloud/synthread/impl.clj#L66

Perhaps we could add to the defrecord abstraction to support well the kind of things that synthread code is doing clumsily, and then walk could take advantage of that.

Comment by Alex Miller [ 27/Feb/14 2:11 PM ]

@Chouser, can you file a new ticket related to this? It's hard to manage work on something from comments on a closed ticket.

Comment by Alex Miller [ 27/Feb/14 3:54 PM ]

@Chouser - Never mind! I was thinking this was the change that went into 1.6. Carry on.

Comment by Nicola Mometto [ 27/Feb/14 5:17 PM ]

Alex, for what it matters clojure-1.6.0 after CLJ-1105 exibits the same behaviour as described by Chouser for this patch

Generated at Thu Jul 28 03:46:29 CDT 2016 using JIRA 4.4#649-r158309.