Did you use a reasonable number of concurrent sessions in your testing?
Testing with just a single session is an easy trap to fall into, and with
things like sequences (Oracle sequences particularly) the effects of
concurrency on performance of inserts and size of index can be dramatic as
you go from single session to even a fairly small number of concurrent
sessions. (The same effect may appear with UUIDs on some platforms
depending on how the UUID is generated.)
Regards
Jonathan Lewis
On Sat, 13 Apr 2024 at 09:37, ahmed.fikri@xxxxxxxxxxx <
ahmed.fikri@xxxxxxxxxxx> wrote:
Hi,
For those who need to decide which primary key type to choose, I conducted
a test myself, initially on PostgreSQL, and I believe the results would be
similar for Oracle.
The test involved creating and saving 1 million records using Java (with
Hibernate) as the client:
strategy Field Type entities creation elapsed time (ms) saving entities
elapsed Time (ms) Table Size
(mb) Index Size
(mb)
UUID String 7768 181184 80 73
UUID UUID 7763 172367 57 37
Sequence Long 10036 163351 49 21
UUIDs consist of two Longs, which means they require double the space
compared to a single Long. However, despite this overhead, we observed that
the creation of entities using UUIDs at the client-side is faster compared
to using sequences. This is due to the fact that UUIDs allow for
client-side ID generation, reducing the need for round trips to the
database during entity creation (Despite efforts to minimize its impact for
sequences, we still observe its effect).
On the other hand, saving entities with Long IDs takes less time than
UUIDs. This is because Longs require less storage space and hence result in
quicker database operations.
In my case, working with microservices and distributed systems, the
preferred primary key type turned out to be UUID
Regards
Ahmed
-----Original-Nachricht-----
Betreff: Re: UUID vs. Sequential ID as Primary
Datum: 2024-04-11T21:36:16+0200
Von: "Jonathan Lewis" <jlewisoracle@xxxxxxxxx>
An: "list, oracle" <oracle-l@xxxxxxxxxxxxx>
When considering the overheads and side effects of sequences it's worth
remembering that in 12c Oracle introduced the "scale" and "extend" options
to prepend the instance id and session id to the generated value so that
contention between instances and between sessions on the same instance
would be minimised.
It's also worth remembering that 19c introduced an automatic resizing
strategy for the sequence cache (which introduced problems for some people,
especially in RAC) to work around the contention at sites that didn't set a
sensible cache size for their sequences. (See comments on this note: Sequence
Accelerator | Oracle Scratchpad (wordpress.com)
<https://jonathanlewis.wordpress.com/2021/08/06/sequence-accelerator/> )
Regards
Jonathan Lewis
On Thu, 11 Apr 2024 at 18:53, ahmed.fikri@xxxxxxxxxxx <
ahmed.fikri@xxxxxxxxxxx> wrote:
Thank you all for your insightful responses. I share the concern about
mistakenly assuming uniqueness, as it could have serious consequences down
the line.
As for Peter's suggestion, the challenge lies in the fact that the client
doesn't handle the insertion process; it simply needs to generate a unique
ID for an entity without directly "*interacting"* with the database.
This can be achieved if the client understands how IDs are generated, such
as being aware of a sequence on the database side and can access that
sequence. However, when using identities, there's a significant hurdle
because the client lacks access to the internally generated sequence, even
if one is utilized server-side. Consequently using IDENTITY leads to poor
performance as just creating a row at client side require round trip to the
database (The client determines when and whether to insert eventual rows
into the database, so it creates a sort of local cache that should, at a
certain point, mirror the database). The system only functions smoothly if
the client can interact with the sequence directly.
Personally, I lean towards using sequences, but I hesitate to recommend
them to others without being able to precisely justify why. Perhaps there's
a benefit to using UUIDs that I'm not yet aware of.
