26

I decided to write a singly-linked list, and had the plan going in to make the internal linked node structure immutable.

I ran into a snag though. Say I have the following linked nodes (from previous add operations):

1 -> 2 -> 3 -> 4

and say I want to append a 5.

To do this, since node 4 is immutable, I need to create a new copy of 4, but replace its next field with a new node containing a 5. The problem is now, 3 is referencing the old 4; the one without the appended 5. Now I need to copy 3, and replace its next field to reference the 4 copy, but now 2 is referencing the old 3...

Or in other words, to do an append, the entire list seems to need to be copied.

My questions:

  • Is my thinking correct? Is there any way to do an append without copying the entire structure?

  • Apparently "Effective Java" contains the reccomendation:

    Classes should be immutable unless there's a very good reason to make them mutable...

    Is this a good case for mutability?

I don't think this is a duplicate of the suggested answer since I'm not talking about the list itself; that obviously has to be mutable to conform to the interface (without doing something like keeping the new list internally and retrieving it via a getter. On second thought though, even that would require some mutation; it would just be kept to a minimum). I'm talking about whether or not the internals of the list must be immutable.

  • I would say yes - the rare immutable collections in Java are either ones with mutating methods overridden to throw, or the incredibly expensive CopyOnWritexxx classes used for multi-threading. Nobody expects collections to be immutable really (although it does create some quirks) – Ordous Aug 12 '15 at 14:45
  • 9
    (Comment on the quote.) That, a great quote, is often greatly misunderstood. Immutability should be applied to ValueObject because of its benefits, but if you are developing in Java, it is impractical to use immutability in places where mutability is anticipated. Most of the time, a class has certain aspects that should be immutable, and certain aspects that needs to be mutable based on requirements. – rwong Aug 12 '15 at 14:55
  • 2
    related (possibly a duplicate): Is collection object better being immutable? "Without performance overheads, can this collection(class LinkedList) be introduced as immutable collection?" – gnat Aug 12 '15 at 18:04
  • Why would you make an immutable linked list? The biggest benefit of a linked list is the mutability, wouldn't you be better off with an array? – Pieter B Aug 13 '15 at 6:36
  • 3
    Could you please help me understand your requirements? You want to have an immutable list, that you want to mutate? Isn't that a contradiction? What do you mean with "the internals" of the list? Do you mean that previously added nodes should not be modifiable later on, but only allow appending to the last node in the list? – null Aug 13 '15 at 13:07
46

With lists in functional languages, you nearly always work with a head and tail, the first element and the remainder of the list. Prepending is much more common because, as you surmised, appending requires copying the entire list (or other lazy data structures that don't precisely resemble a linked list).

In imperative languages, appending is much more common, because it tends to feel more natural semantically, and you don't care about invalidating references to previous versions of the list.

As an example of why prepending doesn't require copying the entire list, consider you have:

2 -> 3 -> 4

Prepending a 1 gives you:

1 -> 2 -> 3 -> 4

But note that it doesn't matter if someone else is still holding a reference to 2 as the head of their list, because the list is immutable and the links only go one way. There's no way to tell the 1 is even there if you only have a reference to 2. Now, if you appended a 5 onto either list, you'd have to make a copy of the entire list, because otherwise it would appear on the other list as well.

  • Ahh, good point on why prepending doesn't require a copy; I didn't think of that. – Carcigenicate Aug 12 '15 at 15:20
17

You are correct, appending requires copying the whole list if you are unwilling to mutate any nodes in-place. For we need to set the next pointer of the (now) second-to-last node, which in an immutable setting creates a new node, and then we need to set the next pointer of the third-to-last node, and so on.

I think the core problem here is not immutability, nor is the append operation ill-advised. Both are perfectly fine in their respective domains. Mixing them is bad: The natural (efficient) interface for an immutable list emphasized manipulation at the front of the list, but for mutable lists it is often more natural to build the list by successively appending the items from first to last.

Therefore I suggest that you make up your mind: Do you want an ephemeral interface or a persistent one? Do most operations produce a new list and leave an unmodified version accessible (persistent), or do you write code like this (ephemeral):

list.append(x);
list.prepend(y);

Both choices are fine, but the implementation should mirror the interface: A persistent data structure benefits from immutable nodes, while an ephemeral one needs internal mutability to actually fulfill the performance promises it implicitly makes. java.util.List and other interfaces are ephemeral: Implementing them on an immutable list is inappropriate and in fact a performance hazard. Good algorithms on mutable data structures are often quite different from good algorithms on immutable data structures, so dressing up an immutable data structure as a mutable one (or vice versa) invites bad algorithms.

While a persistent list has some disadvantages (no efficient appending), this need not be a serious problem when programming functionally: Many algorithms can be formulated efficiently by shifting the mindset and using higher-order functions such as map or fold (to name two relatively primitive ones), or by prepending repeatedly. Furthermore, nobody is forcing you to use only this data structure: When others (ephemeral, or persistent but more sophisticated) are more appropriate, use them. I should also note that persistent lists have some advantages for other workloads: They share their tails, which can conserve memory.

4

If you have a singly linked list you will be working with the front if it more than you would with the back.

Functional languages like prolog and haskel provide easy ways to get the front element and the rest of the array. Appending to the back is a O(n) operation with copies each node.

  • 1
    I've used Haskell. Afaik, it also circumvents part of the problem by using laziness. I'm doing appends because I thought that's what expected by the List interface (I could be wrong there though). I don't think that bit really answers the question either though. The entire list would still need to be copied; it would just make access to the last element added faster since a full traversal wouldn't be required. – Carcigenicate Aug 12 '15 at 14:56
  • Creating a class according to an interface that doesn't match the underlying data structure/algorithm is a invitation to pain and inefficiencies. – ratchet freak Aug 12 '15 at 14:59
  • The JavaDocs explicitly use the word "append". You're saying it's better to ignore that for the sake of implementation? – Carcigenicate Aug 12 '15 at 15:01
  • 2
    @Carcigenicate no I'm saying that it's a mistake to try and fit a singly linked list with immutable nodes into the mold of java.util.List – ratchet freak Aug 12 '15 at 15:02
  • 1
    With laziness there would not be a linked list anymore; there is no list, hence there is no mutation (append). Instead there is some code that generates the next item upon request. But it cannot be used everywhere where a list is used. Namely, if you write a method in Java that expects to mutate a list that is passed in as an argument, that list has to be mutable in the first place. The generator approach requires a complete restructuring (reorganization) of code to make it work - and to make it possible to do away with the list physically. – rwong Aug 12 '15 at 15:03
4

As others have pointed out, you are correct that an immutable singly-linked list requires copying the entire list when performing an append operations.

Often you can use the workaround of implementing your algorithm in terms of cons (prepend) operations, and then reverse the final list once. This still needs to copy the list once, but the complexity overhead is linear in the length of the list, whereas by repeatedly using append you can easily get a quadratic complexity.

Difference lists (see e.g. here) are an interesting alternative. A difference list wraps a list and provides an append operation in constant time. You basically work with the wrapper as long as you need to append, and then convert back to a list when you are finished. This is somehow similar to what you do when you use a StringBuilder to construct a string, and at the end get the result as a String (immutable!) by calling toString. One difference is that a StringBuilder is mutable, but a difference list is immutable. Also, when you convert a difference list back to a list, you still have to construct the entire new list but, again, you only need to do this once.

It should be quite easy to implement an immutable DList class that provides a similar interface as Haskell's Data.DList.

4

You must watch this great video from 2015 React conf by Immutable.js's creator Lee Byron. It will give you pointers and structures to understand how to implement an efficient immutable list which do not duplicate content. The basics ideas are : - as long as two lists use identical nodes (same value, same next node), the same node is used - when the lists start differing, a structure is created at the node of divergence, which holds pointers to the next particular node of each list

This image from a react tutorial may be clearer than my broken English :enter image description here

2

It's not strictly Java, but I suggest you read this article on an immutable, but performant indexed persistent data structure written in Scala:

http://www.codecommit.com/blog/scala/implementing-persistent-vectors-in-scala

Since it's a Scala data structure, it can be used from Java too (with a bit more verbosity). It's based on a data structure available in Clojure, and I'm sure there's a more "native" Java library offering it as well.

Also, a note on constructing immutable data structures: what you usually want is some sort of "Builder", which allows you to "mutate" your "under-construction" data structure by appending elements to it (within a single thread); after you're done appending, you call a method on the "under-construction" object such as .build() or .result(), which "builds" the object, giving you an immutable data structure that you can then safely share.

1

An approach which can sometimes be helpful would be to have two classes of list-holding objects--a forward-list object with a final reference to the first node in a forward-linked list and an initially-null non-final reference which (when non-null) identifies a reverse-list object holding the same items in the opposite order, and a reverse-list object with a final reference to the last item of a reverse-linked list and an initially-null non-final reference which (when non-null) identifies a forward-list object holding the same items in opposite order.

Prepending an item to a forward-list or appending an item to a reverse-list would merely require creating a new node which links to the node identified by the final reference and creating a new list object of the same type as the original, with a final reference to that new node.

Appending an item to a forward-list or prepending to a reverse-list would require having a list of the opposite type; the first time that is done with a particular list it should create a new object of the opposite type and store a reference; repeating the action should reuse the list.

Note that the external state of a list object will be considered the same whether its reference to an opposite-type list is null or identifies an opposite-order list. There would be no need to make any variables final, even when using multi-threaded code, since every list object would have a final reference to a complete copy of its contents.

If code on one thread creates and cache a reversed copy of a list, and the field in which that copy is cached isn't volatile, it would be possible that code on another thread might not see the cached list, but the only adverse effect from that would be that the other thread would need to do extra work building another reversed copy of the list. Because such action would at worst impair efficiency, but would not affect correctness, and because volatile variables create efficiency impairments of their own, it will often be better to have the variable be non-volatile and accept the possibility of occasional redundant operations.

protected by gnat Aug 14 '15 at 21:53

Thank you for your interest in this question. Because it has attracted low-quality or spam answers that had to be removed, posting an answer now requires 10 reputation on this site (the association bonus does not count).

Would you like to answer one of these unanswered questions instead?

Not the answer you're looking for? Browse other questions tagged or ask your own question.