Hash delete followed by insert performance issues

Help & Support
1

To simplify I have an object that implements a rebind(p, q) as an unbind(p) followed by a bind(p, q). It was easier to wire everything up this way and I had better things to do, and still have.

So deeply below there are a bunch of hashes, and unbind and bind operations map to Hash#delete and Hash#[]= on those hashes correspondingly. This way I stumbled upon the following issue:

require "benchmark"

Benchmark.ips do |x|
  h0 = Hash(Int32, Int32).new
  h1 = Hash(Int32, Int32).new(initial_capacity: 10_000_000)
  10_000_000.times { |i| h1[i] = i }

  x.report("0") do
    h0.delete(100)
    h0[100] = 123
  end

  x.report("10000000") do
    h1.delete(100)
    h1[100] = 123
  end
end

On my machine (Ryzen 3 2200G, Crystal 1.12.1, official tarball) this benchmark results in the following:

       0  57.76M ( 17.31ns) (± 9.22%)  0.0B/op           fastest
10000000   2.92k (342.64µs) (±21.05%)  0.0B/op  19789.49× slower

What I expect is for both to take roughly the same time. I don’t understand why this isn’t the case. Maybe my expectations are wrong.

I “fixed” this by writing and using a mutable HAMT, but it’s somewhat slower (like 2.5 times slower on average) no matter how hard I try to optimize it. Maybe because it’s a tree and trees are not cache friendly, maybe because my code is shitty, maybe both, whatever. However, it does give stable performance and did speed up my original… thing.

Mutable HAMT on the same machine:

       0  31.21M ( 32.04ns) (± 1.10%)  0.0B/op        fastest
10000000  18.92M ( 52.84ns) (± 0.37%)  0.0B/op   1.65× slower

So I’m curious why Hash has this issue. Is it a bug? A feature? The CPU being quirky? I really really don’t want to maintain another hash map with its own heap of bugs, one that is (and will be) obviously slower than Crystal’s. I really want to use Hash. The same issue with Set, of course. I had to make my own Set as well. Is there a way to fix this? Or am I understanding/measuring something wrong?

2 Likes
2

Why do you expect this? Both hashes are clearly very different in size. h0 only has a single entries ever (after the first iteration), while h1 has 10M entries. Hash map scales relatively well and it’s even performant for big sizes. But differently sized search spaces still make a significant difference. We’re talking about many orders of magnitude here.
Particularly, Crystal’s Hash implementation has an optimization for very small sizes (<= 8 entries) where it performs a linear scan. With just a single entry that’s an immediate hit.

I have no deeper knowledge of hash map performance, but these measurements seem plausible to me.

3

Actually, the performance degradation happens whenever linear scan is unavailable, i.e. hashes with a capacity as little as 32 elements:

require "benchmark"

Benchmark.ips do |b|
  (2..10).each do |i|
    capacity = 1 << i
    b.report("capacity = #{capacity}") do
      h1 = Hash(Int32, Int32).new(initial_capacity: capacity)
      100_000.times do
        h1.delete(100)
        h1[100] = 123
      end
    end
  end
end

   capacity = 4   1.49k (670.72µs) (± 0.36%)    177B/op         fastest
   capacity = 8   1.49k (671.57µs) (± 0.28%)    176B/op    1.00× slower
  capacity = 16   1.41k (709.23µs) (± 0.32%)    272B/op    1.06× slower
  capacity = 32 193.64  (  5.16ms) (± 0.29%)    595B/op    7.70× slower
  capacity = 64 107.66  (  9.29ms) (± 0.91%)  0.99kB/op   13.85× slower
 capacity = 128  63.62  ( 15.72ms) (± 0.64%)  2.31kB/op   23.44× slower
 capacity = 256  34.58  ( 28.92ms) (± 0.51%)  4.38kB/op   43.12× slower
 capacity = 512  17.73  ( 56.40ms) (± 0.60%)   8.1kB/op   84.09× slower
capacity = 1024   8.96  (111.61ms) (± 0.34%)  16.1kB/op  166.40× slower

And if you do this:

class Hash(K, V)
  def indices_slice
    return nil unless @indices

    case @indices_bytesize
    when 1
      Slice.new(@indices, indices_size)
    when 2
      Slice.new(@indices.as(UInt16*), indices_size)
    else
      Slice.new(@indices.as(UInt32*), indices_size)
    end
  end
end

h1 = Hash(Int32, Int32).new(initial_capacity: 32)
50.times do
  p h1.indices_slice
  h1.delete(100)
  h1[100] = 123
end

you get:

Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 0, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 0, 0, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 0, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 0, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 0, 0, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 0, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 0, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 0, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 0]
Bytes[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]
Bytes[9, 10, 11, 12, 13, 14, 15, 16, 17, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8]

It looks like Hash#@indices is never cleared when deleting a single element.

3 Likes
4

Uhmm is it comparing to dead entries/tombstones or something? Like, skipping over them? Hash colliding with them?

5

I think I agree with indices not being cleared. At least I see a set_index call in upsert and no mention of index or indices in delete_impl and delete_entry_and_update_counts…

6

@entries contains the actual hash-key-value triples, and @indices is an array of indices into those entries. Only when @indices is half-full does #upsert call #resize and #do_compaction.

7

Okay I read the docs :) It says:

We also keep track of the number of deleted entries (@deleted_count). When an
entry is deleted we just mark it as deleted by using the special hash value 0.
Only when the hash needs to be resized we do something with this instance variable:
if we have many deleted entries (at least as many as the number of non-deleted
entries) we compact the map and avoid a resize. Otherwise we remove the non-deleted
entries but also resize both the @entries and @indices buffer. This probably
avoids an edge case where one deletes and inserts an element and there is a constant
shift of the buffer (expensive).

I guess the bold thing is it?

8

Well yeah a set_index(index, -1) (set_index increments by 1 unconditionally) seems to fix the issue

private def delete_impl(key) : Entry(K, V)?
# ...
      # We found a non-empty slot, let's see if the key we have matches
      entry = get_entry(entry_index)
      if entry_matches?(entry, hash, key)
        delete_entry_and_update_counts(entry_index)
        set_index(index, -1) # << HERE
        return entry
      else
        # If it doesn't, check the next index...
        index = next_index(index)
      end
# ...

But I am too unfamiliar with the Hash implementation to foresee the consequences of putting it there… But it fails a hash spec test, “inserts many”. So I guess this won’t work.

9

A proper fix involves defining another special index value to distinguish an empty index (nothing was ever inserted here) from a deleted index (something was here), otherwise Hash would break as it scans @indices linearly for an insertable slot.

10

Okay, I see. “We could use -1 for this but because of an optimization we’ll explain later we use 0, and all other values represent indices which are 1 less than their actual value (so value 3 means index 2).” I thought it had the distinction between -1 and 0 already. The "could"s and the "but"s I skimmed over…

11

I am now trying to do this in Optimize `Hash` for repeated removals and insertions by HertzDevil · Pull Request #14539 · crystal-lang/crystal · GitHub. The problem is that there could still be duplicate entries:

# generates hash collisions whenever two `x`s are congruent modulo twice the hash capacity
record BadHash, x : Int32 do
  def hash
    0x80000000_u32 | @x
  end
end

h = Hash(BadHash, Nil).new(initial_capacity: 32)
h[BadHash.new(0)] = nil
h[BadHash.new(64)] = nil
h.indices_slice # => Bytes[2, 3, 0, ...]

h.delete(BadHash.new(0))
h.indices_slice # => Bytes[1, 3, 0, ...]

h[BadHash.new(64)] = nil
h.indices_slice # => Bytes[4, 3, 0, ...]
h               # => {BadHash(@x=64) => nil, BadHash(@x=64) => nil}

because you cannot simply insert at the first deleted entry; if there were hash collisions, the first collision might occur after the first deleted entry. Yet if we keep searching past the deleted index then nothing is improved. So we need to somehow reorder the indices to ensure that doesn’t happen:

# shifts the 1 two slots to the right
h = Hash(BadHash, Nil).new(initial_capacity: 32)
h[BadHash.new(0)] = nil
h[BadHash.new(64)] = nil
h[BadHash.new(128)] = nil
h.delete(BadHash.new(0))
h.indices_slice # => Bytes[3, 4, 1, 0, ...]

# shifts the 1 one slot to the right
h = Hash(BadHash, Nil).new(initial_capacity: 32)
h[BadHash.new(0)] = nil
h[BadHash.new(64)] = nil
h[BadHash.new(2)] = nil
h.delete(BadHash.new(0))
h.indices_slice # => Bytes[3, 1, 4, 0, ...]

h = Hash(BadHash, Nil).new(initial_capacity: 32)
h[BadHash.new(0)] = nil
h[BadHash.new(64)] = nil
h[BadHash.new(1)] = nil
h.indices_slice # => Bytes[2, 3, 4, 0, ...]
h.delete(BadHash.new(0))
h.indices_slice # => Bytes[3, 4, 1, 0, ...]
2 Likes
12

I did it anyway, and the tradeoff is that deletion used to be O(1) even in case of hash collisions, but now it is O(n).

2 Likes
13

How do other languages (or hash implementations, C has lots) handle this I wonder? :)

14

Well, merged version looks like half of Robin-Hood hashing. May be it is reasonable to make full? But may be not, since it will complicate insertions as well.