ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Roman Kondakov <kondako...@mail.ru.INVALID>
Subject Re: Page eviction for in-memory mode with enabled MVCC
Date Fri, 14 Dec 2018 12:36:37 GMT
Vladimir,

I was thinking about your proposal to not evict locked and recent (the 
transaction that created the record is still active) entries from the 
cache. Let's imagine next situation: we have almost full memory and two 
transactions:

1. txA: "SELECT * FOR UPDATE"

2. txB: "INSERT ...many keys here..."

In this case txA locks all entries in the cache, and therefore we cannot 
evict any of them. If then txB is trying to add a lot of data, it lead 
us to the OOM situation, which user is trying to avoid using cache 
evictions.

I see two ways how to deal with this issue:

1. Allow OOM in MVCC caches with configured evictions and warn user 
about it in the docs.

2. Give up with the repeatable read guaranties in case of evictions for 
MVCC caches and warn users about it in the documentation.

Second variant looks better for me because user may not expect OOM when 
he has configured eviction policy for cache.

What do you think?


-- 
Kind Regards
Roman Kondakov

On 13.12.2018 22:33, Vladimir Ozerov wrote:
> It's hard to believe that entries are not locked on backups, because we
> wrtite data right away. Even if it so, it should be very easy to fix - just
> do not evict and entry if it was created or deleted by currently active
> transaction.
>
> On Thu, Dec 13, 2018 at 10:28 PM Roman Kondakov <kondakov87@mail.ru.invalid>
> wrote:
>
>> Vladimir,
>>
>> we do not lock entries on backups when MVCC is enabled and therefore we
>> don't avoid entry eviction from backup by locking. So, your first
>> scenario with primary stop is still relevant.
>>
>>
>> --
>> Kind Regards
>> Roman Kondakov
>>
>> On 13.12.2018 22:14, Vladimir Ozerov wrote:
>>> No, I mean that we should think about what kind of guarantees it
>> possible.
>>> My proposal was to prevent evictions of locked entries. This way we can
>> say
>>> users: "if you want true REPEATABLE_READ when evictions are enabled, then
>>> make sure to lock entries on every access". This effectively means that
>> all
>>> SELECT's should be replaced with "SELECT FOR UPDATE".
>>>
>>> On Thu, Dec 13, 2018 at 10:09 PM Roman Kondakov
>> <kondakov87@mail.ru.invalid>
>>> wrote:
>>>
>>>> Vladimir,
>>>>
>>>> correct me please if i misunderstood your thought. So, if eviction is
>>>> not about a consistency at all, we may evict keys in any way because
>>>> broken repeatable read semantics is not the biggest problem here. But we
>>>> should add some notes about it to user documentation. Right?
>>>>
>>>>
>>>> --
>>>> Kind Regards
>>>> Roman Kondakov
>>>>
>>>> On 13.12.2018 17:45, Vladimir Ozerov wrote:
>>>>> Roman,
>>>>>
>>>>> I would start with the fact that eviction can never be consistent
>> unless
>>>> it
>>>>> utilizes atomic broadcast protocol, which is not the case for Ignite.
>> In
>>>>> Ignite entries on node are evicted independently.
>>>>>
>>>>> So you may easily get into situation like this:
>>>>> 1) Start a cache with 1 backup and FULL_SYNC mode
>>>>> 2) Put a key to primary node
>>>>> 3) Stop primary
>>>>> 4) Try reading from new primary and get null because key was evicted
>>>>> concurrently
>>>>>
>>>>> Or:
>>>>> 1) Start a transaction in PESSIMISTIC/READ_COMMITTED mode
>>>>> 2) Read a key, get value
>>>>> 3) Read the same key again, get null
>>>>>
>>>>> So in reality the choice is not between consistent and inconsistent
>>>>> behavior, but rather about degree of inconsistency. Any solution is
>>>>> possible as long as we can explain it to the user. E.g. "do not evict
a
>>>> key
>>>>> if it is either write-locked".
>>>>>
>>>>>
>>>>> On Thu, Dec 13, 2018 at 5:19 PM Vladimir Ozerov <vozerov@gridgain.com>
>>>>> wrote:
>>>>>
>>>>>> Andrey,
>>>>>>
>>>>>> We will not be able to cache the whole data set locally, as it
>>>> potentially
>>>>>> lead to OOME. We will have this only as an option and only for non-SQL
>>>>>> updates. Thus, similar semantics is not possible.
>>>>>>
>>>>>> On Thu, Dec 13, 2018 at 4:56 PM Andrey Mashenkov <
>>>>>> andrey.mashenkov@gmail.com> wrote:
>>>>>>
>>>>>>> Roman,
>>>>>>>
>>>>>>> We have a ticket to improve repeatable_read mode [1] via caching
>>>> entries
>>>>>>> locally.
>>>>>>> This should make mvcc transaction repeatable_read semantic similar
to
>>>>>>> non-mvcc Txs
>>>>>>> and allow us to implement eviction in correct way.
>>>>>>>
>>>>>>> Another way is to introduce mvcc shared (read) entry locks and
evict
>>>> only
>>>>>>> entries if no one hold any lock on it,
>>>>>>> but this looks tricky and error prone as your first one as it
may
>> lead
>>>> to
>>>>>>> eviction policy unexpected behavior,
>>>>>>> e.g some versions can be visible while others - no (evicted).
>>>>>>>
>>>>>>> [1] https://issues.apache.org/jira/browse/IGNITE-7371
>>>>>>>
>>>>>>> On Thu, Dec 13, 2018 at 4:34 PM Ilya Kasnacheev <
>>>>>>> ilya.kasnacheev@gmail.com>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> Hello!
>>>>>>>>
>>>>>>>> Is it possible to 'touch' entries read by MVCC transactions
to
>> ensure
>>>>>>> that
>>>>>>>> they are considered recent and therefore are almost never
targeted
>> by
>>>>>>>> eviction?
>>>>>>>>
>>>>>>>> This is 1) with benefits.
>>>>>>>>
>>>>>>>> Regards,
>>>>>>>> --
>>>>>>>> Ilya Kasnacheev
>>>>>>>>
>>>>>>>>
>>>>>>>> чт, 13 дек. 2018 г. в 16:22, Roman Kondakov
>>>> <kondakov87@mail.ru.invalid
>>>>>>>> :
>>>>>>>>
>>>>>>>>> Hi igniters,
>>>>>>>>>
>>>>>>>>> I need your advice with the following issue. When in-memory
cache
>>>>>>>>> reaches it's memory limit, some data may be purged to
avoid OOM
>>>> error.
>>>>>>>>> This process is described in [1]. For MVCC caches this
eviction may
>>>>>>>>> break repeatable read semantics. For example, if transaction
reads
>>>> key
>>>>>>>>> before eviction, this key is visible for it. But if key
is evicted
>>>>>>> some
>>>>>>>>> time later, this key become invisible to anyone, including
our
>>>>>>>>> transaction, which means broken repeatable read semantics.
>>>>>>>>>
>>>>>>>>> Now we see the following solutions of this problem:
>>>>>>>>>
>>>>>>>>> 1. Ignore broken repeatable read semantics. If cache
is set to
>> allow
>>>>>>>>> data eviction, it may lose it's data. This means that
there is no
>>>>>>>>> valuable information stored in cache and occasional repeatable
read
>>>>>>>>> violations can be tolerated.
>>>>>>>>>
>>>>>>>>> 2. Prohibit eviction of MVCC caches at all. For example,
stop
>> writing
>>>>>>> to
>>>>>>>>> caches and throw an appropriate exception in the case
when there is
>>>> no
>>>>>>>>> free space in page memory. Before this exception Ignite
should do
>>>> it's
>>>>>>>>> best to avoid this situation, for example, evict all
non-mvcc
>> caches
>>>>>>> and
>>>>>>>>> run full vacuum to free as much space as possible.
>>>>>>>>>
>>>>>>>>> First approach is bad because it leads to cache consistency
>>>> violation.
>>>>>>>>> Second approach is bad because it's behavior may be unexpected
to
>>>> user
>>>>>>>>> if he has set an eviction policy for cache, but instead
of eviction
>>>>>>>>> Ignite trying to avoid it as much as possible.
>>>>>>>>>
>>>>>>>>> IMO first approach looks better - it is much simpler
to implement
>> and
>>>>>>>>> met user expectations in all points except possible repeatable
read
>>>>>>>>> violations.
>>>>>>>>>
>>>>>>>>> What do you think?
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> [1] https://apacheignite.readme.io/docs/evictions
>>>>>>>>>
>>>>>>>>> --
>>>>>>>>> Kind Regards
>>>>>>>>> Roman Kondakov
>>>>>>>>>
>>>>>>>>>
>>>>>>> --
>>>>>>> Best regards,
>>>>>>> Andrey V. Mashenkov
>>>>>>>

Mime
View raw message