Accessors path invalidation problem

[colinsurprenant]

This problem was originally reported by @guyboertje in #5212. I am copying the invalidation bug sequence here:

> event = LogStash::Event.new
=> #<LogStash::Event:0x25ce435 @metadata_accessors=#<LogStash::Util::Accessors:0x7ea71fc2 @store={}, @lut={}>, @cancelled=false, @data={"@version"=>"1", "@timestamp"=>"2016-04-15T14:19:31.561Z"}, @metadata={}, @accessors=#<LogStash::Util::Accessors:0x7cd5fcf4 @store={"@version"=>"1", "@timestamp"=>"2016-04-15T14:19:31.561Z"}, @lut={}>>
> event["[a][0]"]
=> nil
> event["[a][0]"] = 42
=> 42
> event["[a]"]
=> {"0"=>42}
> event["[a]"] = [42,24]
=> [42, 24]
> event["[a]"]
=> [42, 24]
> event["[a][0]"]
=> 42
> event["[a]"] = [24,42]
=> [24, 42]
> event["[a][0]"]
=> 42
> event["[a][0]"] = {"a"=> 99, "b" => 98}
=> {"a"=>99, "b"=>98}
> event["[a][0]"]
=> {"a"=>99, "b"=>98}
> event["[a]"]
=> [24, 42]
> event["[a][0]"]
=> {"a"=>99, "b"=>98}
> event["[a][0][b]"]
TypeError: can't convert String into Integer
  from org/jruby/RubyFixnum.java:1143:in `[]'

I wrote a spec with this sequence and confirm the problem on both the Ruby and Java Event implementations.

[colinsurprenant]

The following spec sequence is basically the test above.

expect(event["[a][0]"]).to be_nil

expect(event["[a][0]"] = 42).to eq(42)
expect(event["[a][0]"]).to eq(42)
expect(event["[a]"]).to eq({"0" => 42})

expect(event["[a]"] = [42, 24]).to eq([42, 24])
expect(event["[a]"]).to eq([42, 24])

expect(event["[a][0]"]).to eq(42)

expect(event["[a]"] = [24, 42]).to eq([24, 42])
expect(event["[a][0]"]).to eq(24)

expect(event["[a][0]"] = {"a "=> 99, "b" => 98}).to eq({"a "=> 99, "b" => 98})
expect(event["[a][0]"]).to eq({"a "=> 99, "b" => 98})

expect(event["[a]"]).to eq([{"a "=> 99, "b" => 98}, 42])
expect(event["[a][0]"]).to eq({"a "=> 99, "b" => 98})
expect(event["[a][1]"]).to eq(42)
expect(event["[a][0][b]"]).to eq(98)

• it fails as-is using both Ruby and Java Event implementations
• it passes by removing the @lut caching in the Ruby Accessors class

I will see if/how we can invalidate the @lut only when required.