Re: making the backend's json parser work in frontend code - Mailing list pgsql-hackers

> On Jan 22, 2020, at 7:00 PM, Mark Dilger <mark.dilger@enterprisedb.com> wrote:
>
> I have this done in my local repo to the point that I can build frontend tools against the json parser that is now in
src/commonand also run all the check-world tests without failure.  I’m planning to post my work soon, possibly tonight
ifI don’t run out of time, but more likely tomorrow. 

Ok, I finished merging with Robert’s patches.  The attached follow his numbering, with my patches intended to by
appliedafter his. 

I tried not to change his work too much, but I did a bit of refactoring in 0010, as explained in the commit comment.

0011 is just for verifying the linking works ok and the json parser can be invoked from a frontend tool without error —
Idon’t really see the point in committing it. 

I ran some benchmarks for json parsing in the backend both before and after these patches, with very slight changes in
runtime. The setup for the benchmark creates an unlogged table with a single text column and loads rows of json
formattedtext: 

CREATE UNLOGGED TABLE benchmark (
    j text
);
COPY benchmark (j) FROM '/Users/mark.dilger/bench/json.csv’;


FYI:

wc ~/bench/json.csv
     107 34465023 503364244 /Users/mark.dilger/bench/json.csv

The benchmark itself casts the text column to jsonb, as follows:

SELECT jsonb_typeof(j::jsonb) typ, COUNT(*) FROM benchmark GROUP BY typ;

In summary, the times are:

    pristine    patched
    —————    —————
    11.985    12.237
    12.200    11.992
    11.691    11.896
    11.847    11.833
    11.722    11.936

The full output for the runtimes without the patch over five iterations:


CREATE TABLE
COPY 107
  typ   | count
--------+-------
 object |   107
(1 row)


real    0m11.985s
user    0m0.002s
sys    0m0.003s
  typ   | count
--------+-------
 object |   107
(1 row)


real    0m12.200s
user    0m0.002s
sys    0m0.004s
  typ   | count
--------+-------
 object |   107
(1 row)


real    0m11.691s
user    0m0.002s
sys    0m0.003s
  typ   | count
--------+-------
 object |   107
(1 row)


real    0m11.847s
user    0m0.002s
sys    0m0.004s
  typ   | count
--------+-------
 object |   107
(1 row)


real    0m11.722s
user    0m0.002s
sys    0m0.003s


An with the patch, also five iterations:


CREATE TABLE
COPY 107
  typ   | count
--------+-------
 object |   107
(1 row)


real    0m12.237s
user    0m0.002s
sys    0m0.004s
  typ   | count
--------+-------
 object |   107
(1 row)


real    0m11.992s
user    0m0.002s
sys    0m0.004s
  typ   | count
--------+-------
 object |   107
(1 row)


real    0m11.896s
user    0m0.002s
sys    0m0.004s
  typ   | count
--------+-------
 object |   107
(1 row)


real    0m11.833s
user    0m0.002s
sys    0m0.004s
  typ   | count
--------+-------
 object |   107
(1 row)


real    0m11.936s
user    0m0.002s
sys    0m0.004s


—
Mark Dilger
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company