sync: add new Map method LoadAndDelete

[wkonkel]

Problem

It is not currently possible to determine if a value was deleted from a sync.Map. For example:

actual, ok := m.Load("key")
if ok {
  m.Delete("key")
  doSomething(actual)
}

This is not atomic in that another goroutine could Delete the value between Load and Delete and doSomething() would be called twice.

Previous issues #25396 and #23547 suggested modifying Delete signature which would break backwards compatibility.

Solution

Create a new function:

func DeleteWithLoad(key interface{}) (actual interface{}, deleted bool)

If the key is found in the map, this function would return the deleted value and true. If the value is not found, this would return nil, false. The above code would then become:

actual, deleted := m.DeleteWithLoad("key")
if deleted {
  doSomething(actual)
}

[ianlancetaylor]

CC @bcmills

[bcmills]

On a cosmetic note, I would want to name the method Remove rather than DeleteWithLoad.

[bcmills]

I'm curious what the concrete use-cases are. (The proposed operation seems useful in theory, but I wonder whether it actually is in practice.)

Could you give some examples of existing code using sync.Map or a similar API that would benefit from this API?

[wkonkel]

A practical example would be using sync.Map as an in-memory cache above a remote database. Since database insert/delete operations are relatively expensive, we want to only call them as needed. The LoadOrStore function allows us to perform a single insert operation even if many goroutines are trying to insert the same data. But there is no equivalent way to ensure a single operation for delete.

func (self *Cache) Add(key interface{}, value interface{}) {
  _, loaded := self.syncMap.LoadOrStore(key, value)
  if !loaded {
    self.database.Insert(key, value) // expensive operation
  }
}

func (self *Cache) Remove(key interface{}) {
  _, removed := self.syncMap.Remove(key) // proposed function
  if removed {
    self.database.Delete(key) // expensive operation
  }
}

[bcmills]

That example seems racy, though: nothing in sync.Map serializes the calls to Insert and Delete, so you could end up with an execution trace that looks like:

• LoadOrStore → loaded = false
• Remove → removed = true
• database.Delete(key) (a no-op)
• database.Insert(key, value)

Now your cache is out of sync with the underlying database: the cache contains no mapping for key, but the underlying database maps it to value.

I suspect that you'll find this sort of insert/delete race in most concrete examples, and the fix is generally to add some other exclusion mechanism (for example, a sync.Mutex) on the slow path, and to recheck the sync.Map once more once the exclusive lock has been acquired. In that case, the Load and Delete operations on the map no longer need to be atomic (since they're synchronized externally).

[wkonkel]

By that logic, shouldn't LoadOrStore be removed since it enables similar behavior if 2 goroutines ran in parallel:

• [1] LoadOrStore → loaded = false
• [2] Delete
• [1] if loaded { ... }

By the time if loaded { ... } is evaluated, the underlying sync.Map has changed.

Ultimately, I don't think there is an expectation that the loaded return value from LoadOrStore represents the current state of the Map but rather just what happened when LoadOrStore was called. Similarly, the removed return value from the proposed Remove would have the same expected behavior. It doesn't represent current state but rather the result of the function call. If a guarantee of current state is needed for either LoadOrStore or Remove, then sync.Mutex is the only approach.

In my specific use case, deletes happen substantially later than adds, so there is never Store-Delete contention. Instead, I'm trying to prevent hundreds of goroutines from triggering the same Add or Remove query from running. I currently use LoadOrStore to solve for the contentious Add use case, but am forced to use sync.Mutex to solve for contentious Delete.

[ianlancetaylor]

Ping @bcmills

[bcmills]

Ok, I think I can see what you mean. For example, you may have a wave of Store calls, followed by some external synchronization point, followed by a wave of Remove calls, and at that point you're guaranteed that none of the Removes races with a Store.

In that case, I'd be ok with a new method with a signature along the lines of:

func (m *Map) Remove(key interface{}) (prevValue interface{}, removed bool)

[rsc]

Why call it Remove? Then we'd have Delete and Remove, with no obvious distinction between them (but a big difference).

We have LoadOrStore already; why is this not LoadAndDelete?

[bcmills]

Why call it Remove?

For me “delete” carries the connotation of destroying the thing to be deleted, especially “by blotting out, cutting out, or erasing”. The proposed function erases a key or entry from the map, but does not erase the corresponding value; in contrast, a Store stores the value (not just the key), and a Delete does blot out (at least the reference to) the value.

In contrast, “remove” carries the connotation of “[changing] the location, position, station, or residence of” the thing, which is what we are doing with the proposed function: changing the location (of the reference to the value) from the map to the caller.

We have LoadOrStore already; why is this not LoadAndDelete?

I would be ok with naming it LoadAndDelete, although that seems needlessly verbose.

[wkonkel]

My preference would be naming this Remove and also deprecating Delete.

[bcmills]

@wkonkel, I don't think we need to deprecate Delete. It can potentially be more efficient than Remove when the caller does not need the deleted reference.

[rsc]

@bcmills, I admire your precision with the different meanings of delete and remove but I don't think many users will see that distinction.
@wkonkel, this is an unusual operation - there are plenty of times you may just want delete, and it is more efficient as Bryan pointed out.

I retitled this to be about adding LoadAndDelete.

Are there any objections to adding LoadAndDelete?

[rsc]

Rereading the thread, based on this discussion it seems like this is a likely accept.

Leaving open for a week for final comments.

[rsc]

Accepted.

[gopherbot]

Change https://golang.org/cl/205899 mentions this issue: sync: add new Map method LoadAndDelete