Postgres PL/pgSQL function for UUID v7 and a bonus custom UUID v8 to support microsecond precision as well. Read more here: https://datatracker.ietf.org/doc/rfc9562/

A_UUID_v7_for_Postgres.sql

-- Based off IETF draft, https://datatracker.ietf.org/doc/draft-peabody-dispatch-new-uuid-format/

create or replace function uuid_generate_v7()
returns uuid
as $$
begin
  -- use random v4 uuid as starting point (which has the same variant we need)
  -- then overlay timestamp
  -- then set version 7 by flipping the 2 and 1 bit in the version 4 string
  return encode(
    set_bit(
      set_bit(
        overlay(uuid_send(gen_random_uuid())
                placing substring(int8send(floor(extract(epoch from clock_timestamp()) * 1000)::bigint) from 3)
                from 1 for 6
        ),
        52, 1
      ),
      53, 1
    ),
    'hex')::uuid;
end
$$
language plpgsql
volatile;

-- Generate a custom UUID v8 with microsecond precision
create or replace function uuid_generate_v8()
returns uuid
as $$
declare
  timestamp    timestamptz;
  microseconds int;
begin
  timestamp    = clock_timestamp();
  microseconds = (cast(extract(microseconds from timestamp)::int - (floor(extract(milliseconds from timestamp))::int * 1000) as double precision) * 4.096)::int;

  -- use random v4 uuid as starting point (which has the same variant we need)
  -- then overlay timestamp
  -- then set version 8 and add microseconds
  return encode(
    set_byte(
      set_byte(
        overlay(uuid_send(gen_random_uuid())
                placing substring(int8send(floor(extract(epoch from timestamp) * 1000)::bigint) from 3)
                from 1 for 6
        ),
        6, (b'1000' || (microseconds >> 8)::bit(4))::bit(8)::int
      ),
      7, microseconds::bit(8)::int
    ),
    'hex')::uuid;
end
$$
language plpgsql
volatile;

CONTRIBUTORS.txt

PERFORMANCE: Move from pgcrypto to built-in gen_random_uuid():
  Curtis Summers (https://github.com/csummers)

PERFORMANCE: Use set_bit to upgrade v4 to v7, not set_byte:
PERFORMANCE: Reduce local variable use while still being maintainable
  Rolf Timmermans (https://github.com/rolftimmermans)

LICENSE.txt

Copyright 2023 Kyle Hubert <kjmph@users.noreply.github.com> (https://github.com/kjmph)

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Hello @ardabeyazoglu, it is a bit of a subtle answer. Earlier drafts of UUIDv7 contained sub-second precision bits in the format, that an implementation MAY use. UUIDv8 was for all custom usage that was implementation controlled. Current versions of the draft made UUIDv7 only for millisecond precision, and all sub-millisecond precision was moved to UUIDv8 for custom formats. The implementation attached to this draft for UUIDv7 conforms to the current drafts, while the UUIDv8 in this gist conforms to old UUIDv7 with microsecond precision.

The gist you linked to was an old UUIDv7 implementation with microsecond precision. If you want to compare apples to apples, please compare uuid_generate_v8 in this gist to the other implementation for performance analysis.

Note, uuid_generate_v7 in this gist is sorted in your example query. It is generating that many UUIDs per millisecond that they look unordered in the test query. As Rolf indicated.

Thanks @rolftimmermans for answering these questions, thought I would provide more color if this helpful.

Thanks for detailed clarification @kjmph, I also saw the difference after reading the codes carefully.

@kjmph, FYI

According to current RFC (was 4122 now it's RFC 9562 - Universally Unique IDentifiers (UUIDs) ):

uuidv7 may contain:

1. An OPTIONAL sub-millisecond timestamp fraction (12 bits at maximum) as per Section 6.2 (Method 3).

IMHO, it's ok to fill rand_a with 12 bit of microseconds in uuid v7

IMHO, it's ok to fill rand_a with 12 bit of microseconds in uuid v7

uuid_generate_uuidv7_with_us() is wordy, but it's far from the worst function name I've seen. We could have both!

(We could also have both with an optional boolean argument. I suspect that would be a performance hit, but I spend more time in JS than Postgres, so I don't trust my instincts here.)

Hello everyone, thanks for pointing out the uuidv7 changes that specify the optional 12 bits now. However awkward this is, that's what I called uuidv8 in the function above. Back then I thought the uuidv8 name would make sense, since I added custom 12 bits to extend the millisecond precision to microsecond precision and the draft uuidv8 format was specified for custom implementations. Now that the new draft is out, this "custom" extension for a sub-millisecond timestamp fractions is now optional for uuidv7.

So, the above uuid_generate_v8() function generates a format that looks like this layout:

 0                   1                   2                   3
 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                           unix_ts_ms                          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|          unix_ts_ms           |  ver  |      unix_ts_us       |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|var|                        rand_b                             |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                            rand_b                             |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

It happens to be that uuid_generate_uuidv7_with_us() is a perfectly fine name, but I called it uuid_generate_uuidv8() for historical reasons.

And just to be helpful, a reminder for everyone, I'm using time quantization and mapping the microseconds into 12 bits, so it is still monotonic without any wasted bits. It also happens to mean that the microseconds are approximate, because each discrete interval is ~0.244us (or 244140625 femtoseconds for fun!). Regardless of the details, the point is, I'm aware microseconds are 10bits and I'm properly handling that. :) I got you, open source community. [EDIT: To be clear, this is not my idea, it was specified in the original uuidv7 specification, and I just faithfully reproduced it.]

Glad this has been helpful.

Oh, and thanks for the correct link @vchirikov; I'm updating the header to reference the newer RFC 9562. Also, @jamesarosen, I think it is better to not conditionally generate values as this is highly tuned for performance. Use the uuiv8 function directly. I do the following in my tables, for instance:

row_id uuid primary key default uuid_generate_v8()