The messaging service constructor's body does two main things in this
order:
1. it registers the CLIENT_ID verb with rpc.
2. it initializes the scheduling mechanism in charge of locating the
right scheduling group for each verb.
The registration function uses the scheduling mechanism to determine
the scheduling group for the verb.
This commit simply reverses the order of execution.
Fixes#6628
Tenants get their own connections for statement verbs and are further
isolated from each other by different scheduling groups. A tenant is
identified by a scheduling group and a name. When selecting the client
index for a statement verb, we look up the tenant whose scheduling group
matches the current one. This scheduling group is persisted across the
RPC call, using the name to identify the tenant on the remote end, where
a reverse lookup (name -> scheduling group) happens.
Instead of a single scheduling group to be used for all statement verbs,
messaging_service::scheduling_config now contains a list of tenants. The
first among these is the default tenant, the one we use when the current
scheduling group doesn't match that of any configured tenant.
To make this mapping easier, we reshuffle the client index assignment,
such that statement and statement-ack verbs have the idx 2 and 3
respectively, instead of 0 and 3.
The tenant configuration is configured at message service construction
time and cannot be changed after. Adding such capability should be easy
but is not needed for query classification, the current user of the
tenant concept.
Currently two tenants are configured: $user (default tenant) and
$system.
Per-user SLA means we have connection classifications determined dynamically,
as SLAs are added or removed. This means the classification information cannot
be static.
Fix by making it a non-static vector (instead of a static array), allowing it
to be extended. The scheduling group member pointer is replaced by a scheduling
group as a member pointer won't work anymore - we won't have a member to refer
to.
On the client side, we supply an isolation cookie based on the connection index
On the server side, we convert an isolation cookie back to a scheduling_group.
This has two advantages:
- rpc processes the entire connection using the scheduling group, so that code
is also isolated and accounted for
- we can later add per-user connections; the previous approach of looking at the
verb to decide the scheduling_group doesn't help because we don't have a set of
verbs per user
With this, the main group sees <0.1% usage under simple read and write loads.
Move it from a function-local static to a class static variable. We will want
to extend it in two ways:
- add more information per connection index (like the rpc isolation cookie)
- support adding more connections for per-user SLA
As a first step, make it an array of structures and make it accessible to all
of messaging_service.
Commit 75cf255c67 (repair: Ignore keyspace
that is removed in sync_data_using_repair) is not enough to fix the
issue because when the repair master checks if the table is dropped, the
table might not be dropped yet on the repair master.
To fix, the repair master should check if the follower failed the repair
because the table is dropped by checking the error returned from
follower.
With this patch, we would see
WARN 2020-04-14 11:19:00,417 [shard 0] repair - repair id 1 on shard 0
completed successfully, keyspace=ks, ignoring dropped tables={cf}
when the table is dropped during bootstrap.
Tests: update_cluster_layout_tests.py:TestUpdateClusterLayout.simple_add_new_node_while_schema_changes_test
Fixes: #5942
Changes messaging service rpc to use reloadable tls
certificates iff tls is enabled-
Note that this means that the service cannot start
listening at construction time if TLS is active,
and user need to call start_listen_ex to initialize
and actually start the service.
Since "normal" messaging service is actually started
from gms, this route too is made a continuation.
Since 956b092012 (Merge "Repair based node
operation" from Asias), repair is used by other node operations like
bootstrap, decommission and so on.
Send the reason for the repair, so that we can handle the materialized
view update correctly according to the reason of the operation. We want
to trigger the view update only if the repair is used by repair
operation. Otherwise, the view table will be handled twice, 1) when the
view table is synced using repair 2) when the base table is synced using
repair and view table update is triggered.
Fixes#5930Fixes#5998
This removes the need to include reactor.hh, a source of compile
time bloat.
In some places, the call is qualified with seastar:: in order
to resolve ambiguities with a local name.
Includes are adjusted to make everything compile. We end up
having 14 translation units including reactor.hh, primarily for
deprecated things like reactor::at_exit().
Ref #1
The learning stage of PAXOS protocol leaves behind an entry in
system.paxos table with the last learned value (which can be large). In
case not all participants learned it successfully next round on the same
key may complete the learning using this info. But if all nodes learned
the value the entry does not serve useful purpose any longer.
The patch adds another round, "prune", which is executed in background
(limited to 1000 simultaneous instances) and removes the entry in
case all nodes replied successfully to the "learn" round. It uses the
ballot's timestamp to do the deletion, so not to interfere with the
next round. Since deletion happens very close to previous writes it will
likely happen in memtable and will never reach sstable, so that reduces
memtable flush and compaction overhead.
Fixes#5779
Message-Id: <20200330154853.GA31074@scylladb.com>
Since lwt requests are now running on an owning shard there is no longer
a need to invoke cross shard call on paxos_state level. RPC calls may
still arrive to a wrong shard so we need to make cross shard call there.
Introduce a new verb dedicated for receiving and sending hints:
HINT_MUTATION. It is handled on the streaming connection, which is
separate from the one used for handling mutations sent by coordinator
during a write.
The intent of using a separate connection is to increase fariness while
handling hints and user requests - this way, a situation can be avoided
in which one type of requests saturate the connection, negatively
impacting the other one.
Current LWT implementation uses at least three network round trips:
- first, execute PAXOS prepare phase
- second, query the current value of the updated key
- third, propose the change to participating replicas
(there's also learn phase, but we don't wait for it to complete).
The idea behind the optimization implemented by this patch is simple:
piggyback the current value of the updated key on the prepare response
to eliminate one round trip.
To generate less network traffic, only the closest to the coordinator
replica sends data while other participating replicas send digests which
are used to check data consistency.
Note, this patch changes the API of some RPC calls used by PAXOS, but
this should be okay as long as the feature in the early development
stage and marked experimental.
To assess the impact of this optimization on LWT performance, I ran a
simple benchmark that starts a number of concurrent clients each of
which updates its own key (uncontended case) stored in a cluster of
three AWS i3.2xlarge nodes located in the same region (us-west-1) and
measures the aggregate bandwidth and latency. The test uses shard-aware
gocql driver. Here are the results:
latency 99% (ms) bandwidth (rq/s) timeouts (rq/s)
clients before after before after before after
1 2 2 626 637 0 0
5 4 3 2616 2843 0 0
10 3 3 4493 4767 0 0
50 7 7 10567 10833 0 0
100 15 15 12265 12934 0 0
200 48 30 13593 14317 0 0
400 185 60 14796 15549 0 0
600 290 94 14416 15669 0 0
800 568 118 14077 15820 2 0
1000 710 118 13088 15830 9 0
2000 1388 232 13342 15658 85 0
3000 1110 363 13282 15422 233 0
4000 1735 454 13387 15385 329 0
That is, this optimization improves max LWT bandwidth by about 15%
and allows to run 3-4x more clients while maintaining the same level
of system responsiveness.
invoke_on() guarantees that captures object won't be destroyed until the
future returned by the invoked function is resolved so there's no need
to move key, token, proposal for calling paxos_state::*_impl helpers.
Paxos protocol has three stages: prepare, accept, learn. This patch adds
rpc verb for each of those stages. To be term compatible with Cassandra
the patch calls those stages: prepare, propose, commit.
The verbs are:
* DEFINITIONS_UPDATE (push)
* MIGRATION_REQUEST (pull)
Support was added in a backward-compatible way. The push verb, sends
both the old frozen mutation parameter, and the new optional canonical
mutation parameter. It is expected that new nodes will use the latter,
while old nodes will fall-back to the former. The pull verb has a new
optional `options` parameter, which for now contains a single flag:
`remote_supports_canonical_mutation_retval`. This flag, if set, means
that the remote node supports the new canonical mutation return value,
thus the old frozen mutations return value can be left empty.
This patch silences those future discard warnings where it is clear that
discarding the future was actually the intent of the original author,
*and* they did the necessary precautions (handling errors). The patch
also adds some trivial error handling (logging the error) in some
places, which were lacking this, but otherwise look ok. No functional
changes.
Avoid including the lengthy stream_session.hh in messaging_service.
More importantly, fix the build because currently messaging_service.cc
and messaging_service.hh does not include stream_mutation_fragments_cmd.
I am not sure why it builds on my machine. Spotted this when backporting
the "streaming: Send error code from the sender to receiver" to 3.0
branch.
Refs: #4789
In case of error on the sender side, the sender does not propagate the
error to the receiver. The sender will close the stream. As a result,
the receiver will get nullopt from the source in
get_next_mutation_fragment and pass mutation_fragment_opt with no value
to the generating_reader. In turn, the generating_reader generates end
of stream. However, the last element that the generating_reader has
generated can be any type of mutation_fragment. This makes the sstable
that consumes the generating_reader violates the mutation_fragment
stream rule.
To fix, we need to propagate the error. However RPC streaming does not
support propagate the error in the framework. User has to send an error
code explicitly.
Fixes: #4789
get_rpc_client assumes the messaging_service is not stopped. We should check
is_stopping() before we call get_rpc_client.
We do such check in existing code, e.g., send_message and friends. Do
the same check in the newly introduced
make_sink_and_source_for_stream_mutation_fragments() and friends for row
level repair.
Fixes: #4767
Because inet_address was initially hardcoded to
ipv4, its wire format is not very forward compatible.
Since we potentially need to communicate with older version nodes, we
manually define the new serial format for inet_address to be:
ipv4: 4 bytes address
ipv6: 4 bytes marker 0xffffffff (invalid address)
16 bytes data -> address
- REPAIR_GET_ROW_DIFF_WITH_RPC_STREAM
Get repair rows from follower nodes
- REPAIR_PUT_ROW_DIFF_WITH_RPC_STREAM
Put repair rows to follower nodes
- REPAIR_GET_FULL_ROW_HASHES_WITH_RPC_STREAM:
Get full hashes from follower nodes
Before this patch, repair master and followers use their own schema
version at the point repair starts independently. The schemas can be
different due to schema change. Repair uses the schema to serialize
mutation_fragment and deserialize the mutation_fragment received from
peer nodes. Using different schema version to serialize and deserialize
cause undefined behaviour.
To fix, we use the schema the repair master decides for all the repair
nodes involved.
On top of this patch, we could do another step to make sure all nodes
has the latest schema. But let's do it in a separate patch.
Fixes: #4549
Backports: 3.1
Currently, REPAIR_GET_COMBINED_ROW_HASH RPC verb returns only the
repair_hash object. In the future, we will use set reconciliation
algorithm to decode the full row hashes in working row buf. It is useful
to return the number of rows inside working row buf in addition to the
combined row hashes to make sure the decode is successful.
It is also better to use a wrapper class for the verb response so we can
extend the return values later more easily with IDL.
Fixes#4526
Message-Id: <93be47920b523f07179ee17e418760015a142990.1559771344.git.asias@scylladb.com>
Seastar now supports two RPC compression algorithm: the original LZ4 one
and LZ4_FRAGMENTED. The latter uses lz4 stream interface which allows it
to process large messages without fully linearising them. Since, RPC
requests used by Scylla often contain user-provided data that
potentially could be very large, LZ4_FRAGMENTED is a better choice for
the default compression algorithm.
Message-Id: <20190417144318.27701-1-pdziepak@scylladb.com>
Function messaging_service::get_rpc_client() suppose to either return
existing client or create one and return it. The function is suppose to
be atomic, so after checking that requested client does not exist it is
safe to assume emplace() will succeed. But we saw bugs that made the
function to not be atomic. Lets add an assert that will help to catch
such bugs easier if they will happen in the future.
Message-Id: <20190326115741.GX26144@scylladb.com>
Current code captures a reference to rpc::client in a continuation, but
there is no guaranty that the reference will be valid when continuation runs.
Capture shared pointer to rpc::client instead.
Fixes#4350.
Message-Id: <20190314135538.GC21521@scylladb.com>
Replace stdx::optional and stdx::string_view with the C++ std
counterparts.
Some instances of boost::variant were also replaced with std::variant,
namely those that called seastar::visit.
Scylla now requires GCC 8 to compile.
Signed-off-by: Duarte Nunes <duarte@scylladb.com>
Message-Id: <20190108111141.5369-1-duarte@scylladb.com>
"
=== How the the partition level repair works
- The repair master decides which ranges to work on.
- The repair master splits the ranges to sub ranges which contains around 100
partitions.
- The repair master computes the checksum of the 100 partitions and asks the
related peers to compute the checksum of the 100 partitions.
- If the checksum matches, the data in this sub range is synced.
- If the checksum mismatches, repair master fetches the data from all the peers
and sends back the merged data to peers.
=== Major problems with partition level repair
- A mismatch of a single row in any of the 100 partitions causes 100
partitions to be transferred. A single partition can be very large. Not to
mention the size of 100 partitions.
- Checksum (find the mismatch) and streaming (fix the mismatch) will read the
same data twice
=== Row level repair
Row level checksum and synchronization: detect row level mismatch and transfer
only the mismatch
=== How the row level repair works
- To solve the problem of reading data twice
Read the data only once for both checksum and synchronization between nodes.
We work on a small range which contains only a few mega bytes of rows,
We read all the rows within the small range into memory. Find the
mismatch and send the mismatch rows between peers.
We need to find a sync boundary among the nodes which contains only N bytes of
rows.
- To solve the problem of sending unnecessary data.
We need to find the mismatched rows between nodes and only send the delta.
The problem is called set reconciliation problem which is a common problem in
distributed systems.
For example:
Node1 has set1 = {row1, row2, row3}
Node2 has set2 = { row2, row3}
Node3 has set3 = {row1, row2, row4}
To repair:
Node1 fetches nothing from Node2 (set2 - set1), fetches row4 (set3 - set1) from Node3.
Node1 sends row1 and row4 (set1 + set2 + set3 - set2) to Node2
Node1 sends row3 (set1 + set2 + set3 - set3) to Node3.
=== How to implement repair with set reconciliation
- Step A: Negotiate sync boundary
class repair_sync_boundary {
dht::decorated_key pk;
position_in_partition position
}
Reads rows from disk into row buffers until the size is larger than N
bytes. Return the repair_sync_boundary of the last mutation_fragment we
read from disk. The smallest repair_sync_boundary of all nodes is
set as the current_sync_boundary.
- Step B: Get missing rows from peer nodes so that repair master contains all the rows
Request combined hashes from all nodes between last_sync_boundary and
current_sync_boundary. If the combined hashes from all nodes are identical,
data is synced, goto Step A. If not, request the full hashes from peers.
At this point, the repair master knows exactly what rows are missing. Request the
missing rows from peer nodes.
Now, local node contains all the rows.
- Step C: Send missing rows to the peer nodes
Since local node also knows what peer nodes own, it sends the missing rows to
the peer nodes.
=== How the RPC API looks like
- repair_range_start()
Step A:
- request_sync_boundary()
Step B:
- request_combined_row_hashes()
- reqeust_full_row_hashes()
- request_row_diff()
Step C:
- send_row_diff()
- repair_range_stop()
=== Performance evaluation
We created a cluster of 3 Scylla nodes on AWS using i3.xlarge instance. We
created a keyspace with a replication factor of 3 and inserted 1 billion
rows to each of the 3 nodes. Each node has 241 GiB of data.
We tested 3 cases below.
1) 0% synced: one of the node has zero data. The other two nodes have 1 billion identical rows.
Time to repair:
old = 87 min
new = 70 min (rebuild took 50 minutes)
improvement = 19.54%
2) 100% synced: all of the 3 nodes have 1 billion identical rows.
Time to repair:
old = 43 min
new = 24 min
improvement = 44.18%
3) 99.9% synced: each node has 1 billion identical rows and 1 billion * 0.1% distinct rows.
Time to repair:
old: 211 min
new: 44 min
improvement: 79.15%
Bytes sent on wire for repair:
old: tx= 162 GiB, rx = 90 GiB
new: tx= 1.15 GiB, tx = 0.57 GiB
improvement: tx = 99.29%, rx = 99.36%
It is worth noting that row level repair sends and receives exactly the
number of rows needed in theory.
In this test case, repair master needs to receives 2 million rows and
sends 4 million rows. Here are the details: Each node has 1 billion *
0.1% distinct rows, that is 1 million rows. So repair master receives 1
million rows from repair slave 1 and 1 million rows from repair slave 2.
Repair master sends 1 million rows from repair master and 1 million rows
received from repair slave 1 to repair slave 2. Repair master sends
sends 1 million rows from repair master and 1 million rows received from
repair slave 2 to repair slave 1.
In the result, we saw the rows on wire were as expected.
tx_row_nr = 1000505 + 999619 + 1001257 + 998619 (4 shards, the numbers are for each shard) = 4'000'000
rx_row_nr = 500233 + 500235 + 499559 + 499973 (4 shards, the numbers are for each shard) = 2'000'000
Fixes: #3033
Tests: dtests/repair_additional_test.py
"
* 'asias/row_level_repair_v7' of github.com:cloudius-systems/seastar-dev: (51 commits)
repair: Enable row level repair
repair: Add row_level_repair
repair: Add docs for row level repair
repair: Add repair_init_messaging_service_handler
repair: Add repair_meta
repair: Add repair_writer
repair: Add repair_reader
repair: Add repair_row
repair: Add fragment_hasher
repair: Add decorated_key_with_hash
repair: Add get_random_seed
repair: Add get_common_diff_detect_algorithm
repair: Add shard_config
repair: Add suportted_diff_detect_algorithms
repair: Add repair_stats to repair_info
repair: Introduce repair_stats
flat_mutation_reader: Add make_generating_reader
storage_service: Introduce ROW_LEVEL_REPAIR feature
messaging_service: Add RPC verbs for row level repair
repair: Export the repair logger
...
Change the inter-node protocol so we can propagate the view update
backlog from a base replica to the coordinator through the
mutation_done and mutation_failed verbs.
Signed-off-by: Duarte Nunes <duarte@scylladb.com>
Many headers don't really need to include database.hh, the include can
be replaced by forward declarations and/or including the actually needed
headers directly. Some headers don't need this include at all.
Each header was verified to be compilable on its own after the change,
by including it into an empty `.cc` file and compiling it. `.cc` files
that used to get `database.hh` through headers that no longer include it
were changed to include it themselves.
This patch adds the RPC verbs that are needed by the row level repair.
The usage of those verbs are in the following patches.
All the verbs for row level repair are sent by the repair master.
Repair master asks repair slaves to create repair meta objects, a.k.a,
repair_meta object, to store the repair meta data needed by row level
repair algorithm. The repair meta object is identified by the IP address
of the repair master and a uint32 number repair_meta_id chosen by repair
master. When repair master restarts or is out of the cluster, repair
slaves will detect it and remove all existing repair_meta for the repair
master. When repair slave restarts, the existing repair_meta on the
slave will be gone.
The sync boundary used in the verbs is the position_in_partition of the
last mutation_fragment. In each repair round, peers work on
(last_sync_boundary, current_sync_boundary]
