crash: when read big table of 1000000 rows

[yjhjstz]

I wrote a c++ test code:

for (int i = 0; i < festate->rowgroups.size(); i++) {
        elog(INFO, " group %d ", i);
        try {
            festate->reader
            ->RowGroup(i)
            ->ReadTable(festate->indices, &festate->table);
        } catch(const std::exception& e) {
            elog(ERROR,
                 "parquet_fdw : failed to read table %d: %s",
                 festate->row_group, e.what());
        }
    }

it also crashed when i == 32768 .

[zilder]

Hi @yjhjstz,

Can you provide a backtrace? How many rowgroups are there in your file. 32K sounds like pretty unreasonable number of row groups. Is it possible that you share your parquet file?

[yjhjstz]

sorry , I misuse StreamWriter::SetMaxRowGroupSize(1000) according to example.

[zilder]

Can you please anyway send the file that caused crash or the code that produces it? I'm going to reproduce this issue and either fix the bug in parquet_fdw if there is one or add an extra check if there is limitation in libarrow.

[yjhjstz]

1. git clone https://github.com/yjhjstz/parquet_fdw/tree/dev
2. psql run below:

   
   create extension parquet_fdw;
   create server parquet_srv foreign data wrapper parquet_fdw;

CREATE FOREIGN TABLE test (id int , c1 float4[]) SERVER parquet_srv OPTIONS(filename '/Users/jianghuayang/work/fdw/parquet_fdw/data/test.parquet', sorted 'id');

create or replace function gen_float4_arr(int) returns float4[] as $$
select array_agg((random()*100)::float4) from generate_series(1,$1);
$$ language sql strict;

insert into test select id, gen_float4_arr(64) from generate_series(1,1000000) t(id); select * from test;

[yjhjstz]

by the way , you can reproduce it and join optimize the insert routine .

[sdressler]

Is there any update on this issue? I run into a SEGFAULT when running ANALYZE on a big table.

[zilder]

Hi @sdressler, can you send a backtrace?

[sdressler]

@zilder I can, but figured it out eventually and the schema had mismatching types. I can still get you a backtrace if you want and if it helps to make things more stable.

[zilder]

Yes, that would be helpful. Is it possible that you also provide schemes that you used in parquet and in postgres?

[sdressler]

I am going to open a new bug report.