Serializing and deserializing DataFrame keeping column order?

[jvigneron]

Migrating some scripts from JSON3.jl to JSON.jl, I found that JSON.parse() do not restore the column order when writing / parsing a DataFrame :

test_df = DataFrame(a=[1,2,3],b=["b","bb","bbb"],c=[Date(now()),Date(2026,9,7),Date(2000,1,1)],d=rand(3))

df_json = JSON.json(Tables.rowtable(test_df))
df_json3 = JSON3.write(Tables.rowtable(test_df))

df = DataFrame(JSON.parse(df_json)) 
# 3×4 DataFrame
#  Row │ c           b       a      d        
#      │ String      String  Int64  Float64  
# ─────┼─────────────────────────────────────
#    1 │ 2026-09-07  b           1  0.209237
#    2 │ 2026-09-07  bb          2  0.250778
#    3 │ 2000-01-01  bbb         3  0.557584

# => KO, random column order ?

df3 = DataFrame(JSON3.read(df_json3)) # OK
# 3×4 DataFrame
#  Row │ a      b       c           d        
#      │ Int64  String  String      Float64  
# ─────┼─────────────────────────────────────
#    1 │     1  b       2026-09-07  0.209237
#    2 │     2  bb      2026-09-07  0.250778
#    3 │     3  bbb     2000-01-01  0.557584

# => OK, same column order as original dataframe

Is there any solution to keep the initial column order using JSON.jl (without storing the initial order and reordering the columns after) ?

[quinnj]

What you're describing seems to be old, pre-1.0 JSON.jl behavior (which parsed into Dict by default). In 1.x, we have a custom JSON.Object that preserves parse order of keys and doesn't suffer from the same issue. Can you double check your version?

[jvigneron]

My bad, an old unused package (Gadfly) was restricting the update to the lastest version. Works perfectly now. Closing the issue.