Now that we have the new interface to make readers with ranges, we can
simplify the code a lot.
1) Less readers are needed
before: number of ranges of readers
after: smp::count readers at most
2) No foreign_ptr is needed
There is no need to forward to a shard to make the foreign_ptr for
send_info in the first phase and forward to that shard to execute the
send_info in the second phase.
3) No do_with is needed in send_mutations since si now is a
lw_shared_ptr
4) Fix possible user after free of 'si' in do_send_mutations
We need to take a reference of 'si' when sending the mutation with
send_stream_mutation rpc call, otherwise:
msg1 got exception
si->mutations_done.broken()
si is freed
msg2 got exception
si is used again
The issue is introduced in dc50ce0ce5 (streaming: Make the mutation
readers when streaming starts) which is master only, branch 1.5 is not
affected.
Currenlty we make the mutation readers for streaming at different
time point, i.e.,
do_for_each(_ranges.begin(), _ranges.end(), [] (auto range) {
make a mutation reader for this range
read mutations from the reader and send
})
If there are write workload in the background, we will stream extra
data, since the later the reader is made the more data we need to send.
Fix it by making all the readers before starting to stream.
Fixes#1815
Message-Id: <1479341474-1364-2-git-send-email-asias@scylladb.com>
There are places in which we need to use the column family object many
times, with deferring points in between. Because the column family may
have been destroyed in the deferring point, we need to go and find it
again.
If we use lw_shared_ptr, however, we'll be able to at least guarantee
that the object will be alive. Some users will still need to check, if
they want to guarantee that the column family wasn't removed. But others
that only need to make sure we don't access an invalid object will be
able to avoid the cost of re-finding it just fine.
Signed-off-by: Glauber Costa <glauber@scylladb.com>
Message-Id: <722bf49e158da77ff509372c2034e5707706e5bf.1478111467.git.glauber@scylladb.com>
Wrapping ranges are a pain, so we are moving wrap handling to the edges.
Since cql can't generate wrapping ranges, this means thrift and the ring
maintenance code; also range->ring transformations need to merge the first
and last ranges.
Message-Id: <1478105905-31613-1-git-send-email-avi@scylladb.com>
Remove inclusions from header files (primary offender is fb_utilities.hh)
and introduce new messaging_service_fwd.hh to reduce rebuilds when the
messaging service changes.
Message-Id: <1475584615-22836-1-git-send-email-avi@scylladb.com>
"This series improves repair by
1) using less streaming sessions
2) reducing unnecessary streaming traffic
3) fixing a hang during shutdown
See commit log for "repair: Reduce stream_plan usage", "repair: Reduce
unnecessary streaming traffic" and "streaming: Fail streaming sessions
during shutdown" for details.
Tested with repair_additional_test.py."
Using make_streaming_reader for streaming on the sender side, it has
the following advantages:
- streaming, repair will not pollute the row cache on the sender side
any more. Currently, we are risking evicting all the frequently-queried
partitions from the cache when an operation like repair reads entire
sstables and floods the row cache with swathes of cold data from they
read from disk.
- less data will be sent becasue the reader will only return existing
data before the point of the reader is created, plus bounded amount
of writes which arrive later. This helps reducing the streaming time
in the case new data is being inserted all the time while streaming is
in progress. E.g., adding a new node while there is a lot of cql write
workload.
Fixes#382Fixes#1682
We print the following messages even if there is no stream_session with
that peer. It is a bit confusing.
INFO 2016-09-23 08:26:37,254 [shard 0] stream_session - stream_manager:
Close all stream_session with peer = 127.0.0.1 in on_restart
INFO 2016-09-23 08:26:37,287 [shard 0] stream_session - stream_manager:
Close all stream_session with peer = 127.0.0.3 in on_remove
Print only when the streaming session with the peer exists.
Remove clustering_key_filter_factory and clustering_key_filtering_context.
Use partition_slice directly with a static get_ranges method.
Signed-off-by: Piotr Jastrzebski <piotr@scylladb.com>
From Paweł:
This is another episode in the "convert X to streamed mutations" series.
Hashing mutations (mainly for repair) is converted so that it doesn't
need to rebuild whole mutation.
The first part of the series changes the way streamed mutations deal
with range tombstones. Since it is not necessary to make sure we write
disjoint tombstones to sstables there is no need anymore for streamed
mutations to produce disjoint tombstones and, consequently, no need for
range tombstones to be split into range_tombstone_begin and
range_tombstone_end.
The second part is the actual hashing implementation. However, to ensure
that the hash depends only on the contents of the mutation and no the
way it is stored in different data sources range tombstones have to be
made disjoint before they are hashed.
This series also ensures that any changes caused by streamed mutations
to hashing and streaming do not break repair during upgrade.
The receiving side needs to handle fragmented mutations properly so that
isolation guarantees are not broken. If the receiving node may be an old
one do not fragment mutations.
Signed-off-by: Paweł Dziepak <pdziepak@scylladb.com>
Commit 206955e4 "streaming: Reduce memory usage when sending mutations"
moved streaming mutation limiter from do_send_mutations() to
send_mutations(). The reason for that was that send_mutation() did full
mutation copies. That's no longer the case and streaming limiter should
be moved back to do_send_mutation() in order to provide back pressure to
fragment_and_freeze().
Signed-off-by: Paweł Dziepak <pdziepak@scylladb.com>
If mutations are fragmented during streaming a special care must be
taken so that isolation guarantees are not broken.
Mutations received with flag "fragmented" set are applied to a memtable
that is used only by that particular streaming task and the sstables
created by flushing such memtables are not made visible until the task
is complte. Also, in case the streaming fails all data is dropped.
This means that fragmented mutations cannot benefit from coalescing of
writes from multiple streaming plans, hence separate way of handling
them so that there is no loss of performance for small partitions.
Signed-off-by: Paweł Dziepak <pdziepak@scylladb.com>
plan_id is needed to keep track of the origin of mutations so that if
they are fragmented all fragments are made visible at the same time,
when that particular streaming plan_id completes.
Basically, each streaming plan that sends big (fragmented) mutations is
going to have its own memtables and a list of sstables which will get
flushed and made visible when that plan completes (or dropped if it
fails).
Signed-off-by: Paweł Dziepak <pdziepak@scylladb.com>
dtest takes error level log as serious error. It is not a serious error
for streaming to fail to send a verb and fail a streaming session, for
example, the peer node is gone or stopped. Switch to use log level warn
instead of level error.
Fixes repair_additional_test.py:RepairAdditionalTest.repair_kill_3_test
Fixes: #1335
Message-Id: <0149d30044e6e4d80732f1a20cd20593de489fc8.1465979288.git.asias@scylladb.com>
Limit disk bandwidth to 5MB/s to emulate a slow disk:
echo "8:0 5000000" >
/cgroup/blkio/limit/blkio.throttle.write_bps_device
echo "8:0 5000000" >
/cgroup/blkio/limit/blkio.throttle.read_bps_device
Start scylla node 1 with low memory:
scylla -c 1 -m 128M --auto-bootstrap false
Run c-s:
taskset -c 7 cassandra-stress write duration=5m cl=ONE -schema
'replication(factor=1)' -pop seq=1..100000 -rate threads=20
limit=2000/s -node 127.0.0.1
Start scylla node 2 with low memory:
scylla -c 1 -m 128M --auto-bootstrap true
Without this patch, I saw std::bad_alloc during streaming
ERROR 2016-06-01 14:31:00,196 [shard 0] storage_proxy - exception during
mutation write to 127.0.0.1: std::bad_alloc (std::bad_alloc)
...
ERROR 2016-06-01 14:31:10,172 [shard 0] database - failed to move
memtable to cache: std::bad_alloc (std::bad_alloc)
...
To fix:
1. Apply the streaming mutation limiter before we read the mutation into
memory to avoid wasting memory holding the mutation which we can not
send.
2. Reduce the parallelism of sending streaming mutations. Before we send each
range in parallel, after we send each range one by one.
before: nr_vnode * nr_shard * (send_info + cf.make_reader memory usage)
after: nr_shard * (send_info + cf.make_reader memory usage)
We can at least save memory usage by the factor of nr_vnode, 256 by
default.
In my setup, fix 1) alone is not enough, with both fix 1) and 2), I saw
no std::bad_alloc. Also, I did not see streaming bandwidth dropped due
to 2).
In addition, I tested grow_cluster_test.py:GrowClusterTest.test_grow_3_to_4,
as described:
https://github.com/scylladb/scylla/issues/1270#issuecomment-222585375
With this patch, I saw no std::bad_alloc any more.
Fixes: #1270
Message-Id: <7703cf7a9db40e53a87f0f7b5acbb03fff2daf43.1464785542.git.asias@scylladb.com>
- Do nothing in case the session is closed, to prevent we fire up the
timer again
- Print log info when no progress has been made if the time expires, it
is very useful to debug a idle session
- Grab a reference when the keep alive timer is running
Message-Id: <9f2cc3164696905a6a39c0d072a980765d598dfd.1458782956.git.asias@scylladb.com>
A STREAM_MUTATION_DONE message will signal the receiver that the sender
has completed the sending of streams mutations. When the receiver finds
it has zero task to send and zero task to receive, it will finish the
stream_session, and in turn finish the stream_plan if all the
stream_sessions are finished. We should call receive_task_completed only
after the flush finishes so that when stream_plan is finshed all the
data is on disk.
Fixes repair_disjoint_data_test issue with Glauber's "[PATCH v4 0/9] Make
sure repairs do not cripple incoming load" serries
======================================================================
FAIL: repair_disjoint_data_test
(repair_additional_test.RepairAdditionalTest)
----------------------------------------------------------------------
Traceback (most recent call last):
File "scylla-dtest/repair_additional_test.py",
line 102, in repair_disjoint_data_test
self.check_rows_on_node(node1, 3000)
File "scylla-dtest/repair_additional_test.py",
line 33, in check_rows_on_node
self.assertEqual(len(result), rows, len(result))
AssertionError: 2461
Keeping the mutations coming from the streaming process as mutations like any
other have a number of advantages - and that's why we do it.
However, this makes it impossible for Seastar's I/O scheduler to differentiate
between incoming requests from clients, and those who are arriving from peers
in the streaming process.
As a result, if the streaming mutations consume a significant fraction of the
total mutations, and we happen to be using the disk at its limits, we are in no
position to provide any guarantees - defeating the whole purpose of the
scheduler.
To implement that, we'll keep a separate set of memtables that will contain
only streaming mutations. We don't have to do it this way, but doing so
makes life a lot easier. In particular, to write an SSTable, our API requires
(because the filter requires), that a good estimate on the number of partitions
is informed in advance. The partitions also need to be sorted.
We could write mutations directly to disk, but the above conditions couldn't be
met without significant effort. In particular, because mutations can be
arriving from multiple peer nodes, we can't really sort them without keeping a
staging area anyway.
Signed-off-by: Glauber Costa <glauber@scylladb.com>
Streaming has currently one class, that can be used to contain the read
operations being generated by the streaming process. Those reads come from two
places:
- checksums (if doing repair)
- reading mutations to be sent over the wire.
Depending on the amount of data we're dealing with, that can generate a
significant chunk of data, with seconds worth of backlog, and if we need to
have the incoming writes intertwined with those reads, those can take a long
time.
Even if one node is only acting as a receiver, it may still read a lot for the
checksums - if we're talking about repairs, those are coming from the
checksums.
However, in more complicated failure scenarios, it is not hard to imagine a
node that will be both sending and receiving a lot of data.
The best way to guarantee progress on both fronts, is to put both kinds of
operations into different classes.
This patch introduces a new write class, and rename the old read class so it
can have a more meaningful name.
Signed-off-by: Glauber Costa <glauber@scylladb.com>
Below are 3 possible cases in a stream session, after commit
208b7fa7ba (streaming: Simplify session completion logic) We might
close the session before the exchange of the PREPARE_DONE_MESSAGE
message in case 1). To fix, we defer the sending of mutations after
PREPARE_DONE_MESSAGE is sent at the initiator node.
1)
Initiator Follower
tx rx tx rx
1 0 0 1
send prepare
send back prepare
recev prepare
send mutations (close the session before prepare_done msg is sent)
recv mutations (close session before prepare_done msg is received)
send prepare_done
recv prepare_done and send no mutations
2)
Initiator Follower
tx rx tx rx
0 1 1 0
send prepare
send back prepare
recv prepare
nothing to send
send prepare_done
recv prepare_done and send mutations (close session)
recv mutations (close session)
3)
Initiator Follower
tx rx tx rx
1 1 1 1
send prepare
send back prepare
recv prepare
send mutations
recv mutations, can not close session since we have mutations to send
send prepare_done
recv prepare_done and send mutations (close session)
recv mutations (close session)
Message-Id: <d6510b558565db23202164fa491b883ef3796e58.1458634037.git.asias@scylladb.com>
Both the initiator and follower of a stream session knows how many
transfer task and receive task the stream session contains in the
preparation phase. They use the _transfers and _receivers map to track
the tasks, like below:
std::map<UUID, stream_transfer_task> _transfers;
std::map<UUID, stream_receive_task> _receivers;
A stream_transfer_task will send STREAM_MUTATION verb to transfer data
with frozen_mutation, when all the STREAM_MUTATIONs are sent, it will
send STREAM_MUTATION_DONE to tell the peer the stream_transfer_task is
completed and remove the stream_transfer_task from _transfers map. The
peer will remove the corresponding stream_receive_task in _receivers.
We do not really need the COMPLETE_MESSAGE verb to notify the peer we
have completed sending. It makes the session completion logic much
simpler and cleaner if we do not depend on COMPLETE_MESSAGE verb.
However, to be compatible with older version, we always send a
COMPLETE_MESSAGE message and do nothing in the COMPLETE_MESSAGE handler
and replies a ready future even if the stream_session is closed already.
This way, node with older version will get a COMPLETE_MESSAGE message
and manage to send a COMPLETE_MESSAGE message to new node as before.
Message-Id: <1458540564-34277-2-git-send-email-asias@scylladb.com>
