On 2020-12-04, at 03:07:19, Howard Chu <hyc(a)symas.com> wrote:
james anderson wrote:
>
>> On 2020-12-03, at 23:40:18, Howard Chu <hyc(a)symas.com> wrote:
>>
>> james anderson wrote:
>>>
>>>> On 2020-12-03, at 19:17:46, Howard Chu <hyc(a)symas.com> wrote:
>>>>
>>>> Gábor Melis wrote:
>>>>> On Wed, 2 Dec 2020 at 22:50, james anderson
>>>>> <anderson.james.1955(a)gmail.com> wrote:
>>>>>>
>>>>>>
>>>>>>> On 2020-12-02, at 22:53:58, Howard Chu <hyc(a)symas.com>
wrote:
>>>>>>>
>>>>>>> James Anderson wrote:
>>>>>>>> the mdb_env_open documentation includes in its note about
NOTLS, that
>>>>>>>>
>>>>>>>> A read-only transaction may span threads if the user
synchronizes its use.
>>>>>>>>
>>>>>>>> to which read-only operations would this constraint
apply?
>>>>>>>
>>>>>>> It depends.
>>>>>>>
>>>>>>> The only safe approach is to ensure that a txn is not active
simultaneously
>>>>>>> in multiple threads.
>>>>>>
>>>>>> where “active” includes read-only cursors?
>>>>>>
>>>>>> does mean, either one constrains the threads such that there can
be no parallel access to the database, or each thread must establish its own transaction,
in which case there is no guarantee that they operate om the same database state?
>>>>>
>>>>> Chiming in here, a cleaner api could be to allow starting a
>>>>> transaction with a given txn id. That way one would have separate
>>>>> transaction objects, but consistent state. The client code would
need
>>>>> to synchronize threads a bit to guarantee that the txn id is still
>>>>> valid, but this would be more lightweight and easier to reason
about.
>>>>
>>>> In an actively written database there is no legitimate use case for
>>>> opening a new transaction on anything but the newest version of the
>>>> data. Reading or depending on stale data would be an application bug.
>>>
>>> without considering the relation between that notion and the management of
data in a bitemporal store, the question remains, how are two independent threads to
ensure that they are reading the same “newest” version when some other, likewise
independent, process may commit a write transaction in the time interval between the
instants of the respective read transaction begins?
>>
>> How would you do this in any other database system?
>
> i would expect it to permit one of the alternatives which has been mentioned:
> - allow multiple threads to perform read operations in the context of a single
transaction
No. A transaction is a single unit for concurrency control. Allowing multiple threads
to operate within a single transaction means you have no control, and thus invites
memory corruption.
you note in your paper on lmdb that, “because of [mvcc] isolation read accesses ... always
have a self-consistent view of the database."
how are operations performed by threads within a transaction to corrupt this view when
their access is through read-only memory and they do nothing to change it?
No transaction system in existence supports this.
> - allow each thread to create a sub-transaction from an initial parent transaction
and then operate on its child transaction
Same as above. The docs on child transactions are quite clear - when a child transaction
is active, the parent transaction cannot be used again until the child finishes.
> - allow a thread to specify the revision identifier of an open transaction as the
state for which it opens its own transaction.
> i would expect to be able to do either of the first two options in an in-mery
database which supports mvcc.
Clearly that would be a broken design.
> i would expect the third option to be available in any database which supports access
to revisions/versions.
That is not what LMDB does. Please read the LMDB design spec.
https://openldap.org/pub/
i have re-read
https://openldap.org/pub/hyc/mdb-paper.pdf with particular attention to
whether your intent or the described implementation decisions necessarily prescribe an
access pattern in which two autonomous threads are guaranteed read access to the same
database state.
as it were, if two threads should happen to begin their respective transaction for
identical revisions, i would expect this to be the case.
what have i overlooked in that exposition which precludes a mechanism which supports
this?
> i would expect the first option to apply to blueprints.
> i would expect the third option to apply to oracle or db2 “versioning”.
---
james anderson | james(a)dydra.com |
http://dydra.com