Re: Potential G2-item cycles under serializable isolation

Hi,

On 2020-06-11 17:40:55 +1200, Thomas Munro wrote:
> +   <para>
> +    The Repeatable Read isolation level is implemented using a technique
> +    known in academic database literature and in some other database products
> +    as <firstterm>Snapshot Isolation</firstterm>.  Differences in behavior
> +    may be observed when compared with systems using other implementation
> +    techniques.  For a full treatment, please see
> +    <xref linkend="berenson95"/>.
> +   </para>

Could it be worthwhile to narrow the "differences in behaviour" bit to
read-write transactions? IME the biggest reason people explicitly use RR
over RC is to avoid phantom reads in read-only transactions. Seems nicer
to not force users to read an academic paper to figure that out?


> @@ -1726,6 +1744,13 @@ SELECT pg_advisory_lock(q.id) FROM
>      see a transient state that is inconsistent with any serial execution
>      of the transactions on the master.
>     </para>
> +
> +   <para>
> +    Access to the system catalogs is not done using the isolation level
> +    of the current transaction.  This has the effect that newly created
> +    database objects such as tables become visible to concurrent Repeatable
> +    Read and Serializable transactions, even though their contents does not.
> +   </para>
>    </sect1>

Should it be pointed out that that's about *internal* accesses to
catalogs? This could be understood to apply to a SELECT * FROM pg_class;
where it actually doesn't apply?

Greetings,

Andres Freund