Once upon a time I got a task to merge duplicate URLs in our production database. It turned out that there were a lot of same urls like and we wanted to merge. Code to normalize URL was easy to develop. Code to migrate the database looked easy as weel. Until I turned the transaction on. Then very cryptic message appeared

 pq: unexpected Parse response 'C'

The error message was telling me nothing. And what is worse ducking did not revealed anything. Except notes about Postgres protocol internals., which was something I did not wanted to dig into. It was a hint from @karolhrdina who have explained me the root cause well enough, so I can share my experience in my blog. Lessons learned - always work with a great colleagues, you can learn a lot from them!

Following text assume reader know about Go, particularly database/sql and lib/pq, Postgres or SQL databases in general.

The code

Following snippet omits a lot of details, but shows the algorithm

rows, err := tx.Query("SELECT id, url, visists FROM schema.url WHERE url LIKE $1 ORDER BY id", pattern)

for rows.Next() {
        err = rows.Scan(&id, &url, &visits)
        urlNormalized := URLNormalize(url)
        if url != urlNormalized {
                _, err = tx.Exec("INSERT INTO schema.url AS u (url, visits) VALUES ($1, 1) ON CONFLICT url DO UPDATE SET visists=u.visists+$2;", urlNormalized, vists)
                _, err = tx.Exec("DELETE FROM schema.url WHERE id=$1", id)
  1. For all urls matching the bad pattern do
  2. Try to normalize and if there is normalized version then
  3. UPSERT new entry or increase a number of visits
  4. … and program crashes

The problem I was facing is connected with how Postgres protocol works. It turns out that you can’t call Exec if you did not read all the data from previous Query. The problem is that most of Postgres drivers including its ownlibpq does fetch all the rows by default. It gives the false impression that following code is legal and works. Go driver lib/pq does not do it. plpgsql does handle this case well. Other language bindings fetch the data by default.

However fetching the data was not an option as there are millions of entries. And using different language than Go was impractical. Url normalization has been written in Go and there are huge differences between language parsers of URLs between languages. For example PHP handles query strings in really bizzare way incompatible with CGI BIN, Perl and any other language. One can build Go code as (C like) shared library, but that would be somewhat big effort.

Fetching data was an option, however given the number of entries, one would need to play with LIMIT and OFFSET to achieve reasonable size of input data.

Do it as plpgsql

There is object called cursor. Cursors are intended exactly for this use case. They encapsulates the query and one can read a few results at the time. And plpgsql uses cursors under the hood. It turns out CURSOR can be used from Go code as well

So Go code changes a bit. There is no *sql.Rows and no Next() method to be used in the loop.

_, err = tx.Exec("DECLARE url_cur CURSOR FOR SELECT id, url, visits FROM schema.url WHERE url LIKE $1 ORDER BY id", pattern)
defer tx.Exec("CLOSE url_cur")

for {
        err = tx.QueryRow("FETCH NEXT FROM url_cur").Scan(&id, &url, &visits)
	if err == sql.ErrNoRows {
	} else if err != nil {
	        _ = tx.Rollback()
        // call tx.Exec with UPSERT and DELETE

The advantages are

  1. No new language in the stack, neither complicated integration of Go code into another language
  2. It is efficient and works with one entry at the time
  3. There are minimal changes to Go code
  4. It works, data are written on tx.Commit()

Big thanks comes to @hypnoglow and his comment at Github issue#635

Logo by kubina@Flickr: []