3-10x slowdown in parsing from JSON3.jl to v1

[jakobnissen]

I'm seeing a 10x slowdown on first JSON3 parse, and a 3x on subsequent parses when comparing JSON.jl v1 (commit ecda14d) with JSON3 v1.14.3. That is not ideal, since parsing JSON is a bottleneck in my program.

To reproduce:

• Make an environment with JSON.jl commit ecda14d and JSON3.jl v1.14.3
• Download the attached gzip file, and decompress it in your directory
• Run the script below, switching the boolean @static if false between runs (I used Julia 1.12.0-rc1)

Timings on my computer:

JSON3:

first_time = 0.408839696
mean_time = 0.20233652939999996

JSON v1

first_time = 4.310216864
mean_time = 0.6266064803999999

File

reference.json.gz

Script

const GENOMES_JSON_T = Vector{Tuple{String, Int, Vector{Tuple{String, Int}}}}
const SEQUENCES_JSON_T = Vector{Tuple{String, Int, Vector{Tuple{String, Int, Int}}}}
const TAXMAPS_JSON_T = Vector{Vector{Tuple{String, String}}}

struct ReferenceJSON
    version::Int
    genomes::GENOMES_JSON_T
    sequences::SEQUENCES_JSON_T
    taxmaps::TAXMAPS_JSON_T
end

s = read("reference.json", String)

f = @static if true # change to switch JSON versions
    using JSON
    JSON.parse
else
    using JSON3
    JSON3.parse
end

first_time = (@timed f(s, ReferenceJSON)).time
@show first_time

y = 0.0
for _ in 1:10
    global y += (@timed f(s, ReferenceJSON)).time
end

mean_time = y / 10
@show mean_time

[quinnj]

Thanks for the report! I'm digging into it. The tuple cases certainly weren't as thoroughly considered in my benchmarking, so they could definitely use a little love.