Date: Fri, 12 Apr 2024 14:19:28 -0500
From: Nathan Bossart <nathandbossart@gmail.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: pgsql-hackers@lists.postgresql.org, alex work <alexwork033@gmail.com>,
	Robert Haas <robertmhaas@gmail.com>
Subject: Re: further improving roles_is_member_of()
Message-ID: <20240412191928.GA2519977@nathanxps13>
References: <20240412041633.GA2315447@nathanxps13>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20240412041633.GA2315447@nathanxps13>
Archived-At: <https://www.postgresql.org/message-id/20240412191928.GA2519977%40nathanxps13>
Precedence: bulk

On Thu, Apr 11, 2024 at 11:16:33PM -0500, Nathan Bossart wrote:
> As shown in the attached work-in-progress patch, this actually ends up
> removing more code than it adds, and it seems to provide similar results to
> HEAD (using the benchmark from the previous thread [0]).  I intend to test
> it with many more roles to see if it's better in more extreme cases.

Even with 100K roles, the Bloom filter added in commit d365ae7 seems to do
a pretty good job at keeping up with the hash table approach.  The callers
of roles_is_member_of() that do list_member_oid() on the returned list
might benefit a little from a hash table, but I'm skeptical it makes much
difference in practice.  This was an interesting test, but I'll likely
withdraw this patch shortly.

-- 
Nathan Bossart
Amazon Web Services: https://aws.amazon.com