This is opened after noticing description is much slower in 3.x. I'm pretty unfamiliar with psycopg (and even more so with cython) so this might be more of a starting point for discussion.

Changes

• Fetching the encoding (even though it's in a Python LRU cahce) was kind of expensive. Now it's cached in the cursor once its fetched once until a reset().
• Cythonified the Column class so libpq can be called directly.
• Made the work to get names + types as lazy as possible.

Benchmarking

Benchmark code (table has ~20 columns):

def benchmark_description(conn, iterations=10000):
    """Benchmark cursor.description access and return elapsed nanoseconds"""
    start_ns = time.perf_counter_ns()
    cur = conn.cursor()
    cur.execute("SELECT * FROM benchmark_table LIMIT 1")

    # Access description many times
    for _ in range(iterations):
        desc = cur.description

    cur.close()
    return time.perf_counter_ns() - start_ns

Rough timing from my machine:

• psycopg2 (2.9.11) - 0.004170051 seconds (4170051 ns)
• psycopg[c] (3.3.2) - 2.072374879 seconds (2072374879 ns)
• psycopg[c] (with this PR) - 0.395435509 seconds (395435509 ns)

Testing

There are no API changes so the existing tests should be sufficient.