Re: [HACKERS] WIP aPatch: Pgbench Serialization and deadlock errors

On Sun, 20 Mar 2022 16:11:43 +0900 (JST)
Tatsuo Ishii <ishii(at)sraoss(dot)co(dot)jp> wrote:

> > Hi Yugo,
> >
> > I tested with serialization error scenario by setting:
> > default_transaction_isolation = 'repeatable read'
> > The result was:
> >
> > $ pgbench -t 10 -c 10 --max-tries=10 test
> > transaction type: <builtin: TPC-B (sort of)>
> > scaling factor: 10
> > query mode: simple
> > number of clients: 10
> > number of threads: 1
> > maximum number of tries: 10
> > number of transactions per client: 10
> > number of transactions actually processed: 100/100
> > number of failed transactions: 0 (0.000%)
> > number of transactions retried: 35 (35.000%)
> > total number of retries: 74
> > latency average = 5.306 ms
> > initial connection time = 15.575 ms
> > tps = 1884.516810 (without initial connection time)
> >
> > I had hard time to understand what those numbers mean:
> > number of transactions retried: 35 (35.000%)
> > total number of retries: 74
> >
> > It seems "total number of retries" matches with the number of ERRORs
> > reported in PostgreSQL. Good. What I am not sure is "number of
> > transactions retried". What does this mean?
>
> Oh, ok. I see it now. It turned out that "number of transactions
> retried" does not actually means the number of transactions
> rtried. Suppose pgbench exectutes following in a session:
>
> BEGIN; -- transaction A starts
> :
> (ERROR)
> ROLLBACK; -- transaction A aborts
>
> (retry)
>
> BEGIN; -- transaction B starts
> :
> (ERROR)
> ROLLBACK; -- transaction B aborts
>
> (retry)
>
> BEGIN; -- transaction C starts
> :
> END; -- finally succeeds
>
> In this case "total number of retries:" = 2 and "number of
> transactions retried:" = 1. In this patch transactions A, B and C are
> regarded as "same" transaction, so the retried transaction count
> becomes 1. But it's confusing to use the language "transaction" here
> because A, B and C are different transactions. I would think it's
> better to use different language instead of "transaction", something
> like "cycle"? i.e.
>
> number of cycles retried: 35 (35.000%)

In the original patch by Marina Polyakova it was "number of retried",
but I changed it to "number of transactions retried" is because I felt
it was confusing with "number of retries". I chose the word "transaction"
because a transaction ends in any one of successful commit , skipped, or
failure, after possible retries.

Well, I agree with that it is somewhat confusing wording. If we can find
nice word to resolve the confusion, I don't mind if we change the word.
Maybe, we can use "executions" as well as "cycles". However, I am not sure
that the situation is improved by using such word because what such word
exactly means seems to be still unclear for users.

Another idea is instead reporting only "the number of successfully
retried transactions" that does not include "failed transactions",
that is, transactions failed after retries, like this;

number of transactions actually processed: 100/100
number of failed transactions: 0 (0.000%)
number of successfully retried transactions: 35 (35.000%)
total number of retries: 74

The meaning is clear and there seems to be no confusion.

Regards,
Yugo Nagata

--
Yugo NAGATA <nagata(at)sraoss(dot)co(dot)jp>