Vladimir,
The query protocol will be changed in both solution.
The tables versions must be added to the 'GridQueryNextPageResponse'
message
and must be compared on the reduce node too. Because the DLL / DML race
may be happens on different nodes.
21.11.2018 15:24, Vladimir Ozerov пишет:
Taras,
Thank you for analysis. I'd like to add that there is another solution -
PME :-) In this case DDL will be blocked until current SELECTs and updates
are finished, and do not allow new operations to be executed. This approach
solves all issues - it avoids deadlocks between query threads and DDL
threads when they are reordered on different nodes, it avoids starvation of
DDL operations, and it doesn't cancel any queries. But there is serious
drawback - performance. The drawback is that it is more complex to
implement (query protocol changes might be required), it blocks the cluster
even when it is needed, and it may destabilize PME mechanism, which is
already on his last legs.
For this reason killing queries which interleave with DDL looks like a
balanced solution for now - it is reasonably simple, allows us to we avoid
OOME in many cases, and do not introduce any additional complexity for
users, as they are already prepared for re-tries.
But I would like to stress one thing - we will need integration with PME at
some point in time anyway. Some DDL operations are blocking in their nature
(e.g. DROP COLUMN). Other DDL operations may be non-blocking, but blocking
implementation may give them serious performance benefits (e.g. CREATE
INDEX).
So I propose to go with your solution for now, and start thinking about SQL
-> PME integration in the background.
Vladimir.
On Wed, Nov 21, 2018 at 2:23 PM Taras Ledkov <[email protected]> wrote:
Hi community,
We will enhance lazy mode for SQL query execution.
Lazy mode overview:
Lazy mode is related to H2 lazy mode when the all query results are not
copied to the RAM in some cases.
The mode it is applicable for SELECTs that doesn't not require
materialize all results in memory, e.g. simple scan plans, IDX lookup,
merge join etc.
And not applicable for SORT by not indexed fields, aggregates, nested
loops joins etc.
When mode is applicable it produces result with iterator-like behavior
on server side and not consume RAM.
So the huge result set may be selected without OOME.
The current implementation.
The current implementation is start separate thread for each query with
'lazy=true'.
This is caused by the our implementation of 'GridH2Table'. In details:
the table's locks.
The table must be locked while result set is completed.
When lazy is disabled a complete result is generated on the first step
of a query execution (then tables unlock)
and result is stored on the node and sent to other node (or client) page
by page.
When lazy is enabled tables are locked until result set delivery to client.
The start new thread causes overhead for requests that returns small
result set.
But current table lock is used `ReentrantReadWriteLock` and we cannot
lock tables from one thread
of QUERY thread pool and unlock in the other thread (when query is
complete or cancel).
The trivial solve is using the 'StampedLock' it solve the lock behavior,
but not solve the table DDL starvation / deadlock.
Example:
Lets the QUERY thread pool contains only one thread. The case is scaled
for any thread pool size.
Write operation that require to exclusive table lock is DDL operation.
1. The query Q0 acquires the shared lock for the table T, send first
page result and leave thread 'threadQP0' control.
2. DDL0 blocks on write lock the table T at the 'threadWP0 '
3. The query Q1 blocks on read lock the 'threadQP0' (because the writer
in the queue).
The deadlock happens. Q0 never can finish and unlock because query pool
hasn't free thread.
The possible solution:
1. Don't use readlock at all. The lock is used only for write /
exclusive (DDL) operations.
2. The DDL (exclusive) operation change the table version.
3. Each read operation (query execution, result page fetch) store the
table version before start and compare with the table version on the
end. If the version is changed the special retry exception is thrown.
CONS:
- The retry logic is less user-friendly. But the distributed SQL cannot
protect the user from implement retry logic totally: e.g. cluster
topology change must handled on user side by retry query implemented by
user, because some data have been delivered to user and we don't track
which data is delivered.
PROS:
- no deadlocks;
- no contention on table lock for SQL query.
What do you think?
--
Taras Ledkov
Mail-To: [email protected]
--
Taras Ledkov
Mail-To: [email protected]
