Fix collect on caches that other tasks modify

src/containers/interface.jl

@@ -133,6 +133,8 @@ Base.isempty(c::AbstractCache) = @lock c.lock isempty(c.index)
 
 # Iteration walks a snapshot copied under the lock, so it is safe while other tasks use the
 # cache (holding the lock across `iterate` calls would deadlock on an early `break`).
+# The length may change between `length` and `iterate`, so `collect` must not rely on it.
+Base.IteratorSize(::Type{<:AbstractCache}) = Base.SizeUnknown()
 function Base.iterate(c::AbstractCache)
     snapshot = @lock c.lock _pairs(c)
     return iterate(c, (snapshot, 1))

test/test_containers.jl

@@ -105,6 +105,24 @@ end
     @test all(((k, v),) -> v == 3k, c)
 end
 
+@testset "$C: collect while other tasks write" for C in CACHETYPES
+    c = C{Int, Int}(; maxsize = 50)
+    done = Threads.Atomic{Bool}(false)
+    writer = Threads.@spawn while !done[]
+        foreach(k -> get!(() -> k, c, k), 1:200)
+        empty!(c)
+        yield()
+    end
+    ok = true
+    for _ in 1:2_000
+        ok &= all(((k, v),) -> k == v, collect(c)) # used to throw when the length changed
+        yield()
+    end
+    done[] = true
+    wait(writer)
+    @test ok
+end
+
 @testset "$C: freed slots release their values" for C in CACHETYPES
     c = C{Int, Base.RefValue{Int}}(; maxsize = 3)
     # build the value inside a function so no local keeps it alive