Seastar is an external library from Scylla's point of view so
we should use the angle bracket #include style. Most of the source
follows this, this patch fixes a few stragglers.
Also fix cases of #include which reached out to seastar's directory
tree directly, via #include "seastar/include/sesatar/..." to
just refer to <seastar/...>.
Closes#10433
Instead of lengthy blurbs, switch to single-line, machine-readable
standardized (https://spdx.dev) license identifiers. The Linux kernel
switched long ago, so there is strong precedent.
Three cases are handled: AGPL-only, Apache-only, and dual licensed.
For the latter case, I chose (AGPL-3.0-or-later and Apache-2.0),
reasoning that our changes are extensive enough to apply our license.
The changes we applied mechanically with a script, except to
licenses/README.md.
Closes#9937
Move replica-oriented classes to the replica namespace. The main
classes moved are ::database, ::keyspace, and ::table, but a few
ancillary classes are also moved. There are certainly classes that
should be moved but aren't (like distributed_loader) but we have
to start somewhere.
References are adjusted treewide. In many cases, it is obvious that
a call site should not access the replica (but the data_dictionary
instead), but that is left for separate work.
scylla-gdb.py is adjusted to look for both the new and old names.
The database, keyspace, and table classes represent the replica-only
part of the objects after which they are named. Reading from a table
doesn't give you the full data, just the replica's view, and it is not
consistent since reconciliation is applied on the coordinator.
As a first step in acknowledging this, move the related files to
a replica/ subdirectory.
Instead of calling get_local_storage_proxy in paxos_state, get it from the
caller (who is, in fact, storage_proxy or one of its components).
Some of the callers, although they are storage_proxy components, don't
have a storage_proxy reference handy and so they ignomiously call
get_local_storage_proxy() themselves. This will be adjusted later.
The other callers who are, in fact, storage_proxy, have to take special
care not to cross a shard boundary. When they do, smp::submit_to()
is converted to sharded::invoke_on() in order to get the correct local instance.
Test: unit (dev)
Closes#9824
Some .cc files over the code include the storage service
for no real need. Drop the header and include (in some)
what's really needed.
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
The write paths in storage_proxy pass replica sets as
std::unordered_set<gms::inet_address>. This is a complex type, with
N+1 allocations for N members, so we change it to a small_vector (via
inet_address_vector_replica_set) which requires just one allocation, and
even zero when up to three replicas are used.
This change is more nuanced than the corresponding change to the read path
abe3d7d7 ("Merge 'storage_proxy: use small_vector for vectors of
inet_address' from Avi Kivity"), for two reasons:
- there is a quadratic algorithm in
abstract_write_response_handler::response(): it searches for a replica
and erases it. Since this happens for every replica, it happens N^2/2
times.
- replica sets for writes always include all datacenters, while reads
usually involve just one datacenter.
So, a write to a keyspace that has 5 datacenters will invoke 15*(15-1)/2
=105 compares.
We could remove this by sending the index of the replica in the replica
set to the replica and ask it to include the index in the response, but
I think that this is unnecessary. Those 105 compares need to be only
105/15 = 7 times cheaper than the corresponding unordered_set operation,
which they surely will. Handling a response after a cross-datacenter round
trip surely involves L3 cache misses, and a small_vector reduces these
to a minimum compared to an unordered_set with its bucket table, linked
list walking and managent, and table rehashing.
Tests using perf_simple_query --write --smp 1 --operations-per-shard 1000000
--task-quota-ms show two allocations removed (as expected) and a nice
reduction in instructions executed.
before: median 204842.54 tps ( 54.2 allocs/op, 13.2 tasks/op, 49890 insns/op)
after: median 206077.65 tps ( 52.2 allocs/op, 13.2 tasks/op, 49138 insns/op)
Closes#8847
A follow up for the patch for #7611. This change was requested
during review and moved out of #7611 to reduce its scope.
The patch switches UUID_gen API from using plain integers to
hold time units to units from std::chrono.
For one, we plan to switch the entire code base to std::chrono units,
to ensure type safety. Secondly, using std::chrono units allows to
increase code reuse with template metaprogramming and remove a few
of UUID_gen functions that beceme redundant as a result.
* switch get_time_UUID(), unix_timestamp(), get_time_UUID_raw(), switch
min_time_UUID(), max_time_UUID(), create_time_safe() to
std::chrono
* remove unused variant of from_unix_timestamp()
* remove unused get_time_UUID_bytes(), create_time_unsafe(),
redundant get_adjusted_timestamp()
* inline get_raw_UUID_bytes()
* collapse to similar implementations of get_time_UUID()
* switch internal constants to std::chrono
* remove unnecessary unique_ptr from UUID_gen::_instance
Message-Id: <20210406130152.3237914-2-kostja@scylladb.com>
This patch introduces a new system table: `system.scylla_table_schema_history`,
which is used to keep track of column mappings for obsolete table
schema versions (i.e. schema becomes obsolete when it's being changed
by means of `CREATE TABLE` or `ALTER TABLE` DDL operations).
It is populated automatically when a new schema version is being
pulled from a remote in get_schema_definition() at migration_manager.cc
and also when schema change is being propagated to system schema tables
in do_merge_schema() at schema_tables.cc.
The data referring to the most recent table schema version is always
present. Other entries are garbage-collected when the corresponding
table schema version is obsoleted (they will be updated with a TTL equal
to `DEFAULT_GC_GRACE_SECONDS` on `ALTER TABLE`).
In case we failed to persist column mapping after a schema change,
missing entries will be recreated on node boot.
Later, the information from this table is used in `paxos_state::learn`
callback in case we have a mismatch between the most recent schema
version and the one that is stored inside the `frozen_mutation`
for the accepted proposal.
Such situation may arise under following circumstances:
1. The previous LWT operation crashed on the "accept" stage,
leaving behind a stale accepted proposal, which waits to be
repaired.
2. The table affected by LWT operation is being altered, so that
schema version is now different. Stored proposal now references
obsolete schema.
3. LWT query is retried, so that Scylla tries to repair the
unfinished Paxos round and apply the mutation in the learn stage.
When such mismatch happens, prior to that patch the stored
`frozen_mutation` is able to be applied only if we are lucky enough
and column_mapping in the mutation is "compatible" with the new
table schema.
It wouldn't work if, for example, the columns are reordered, or
some columns, which are referenced by an LWT query, are dropped.
With this patch we try to look up the column mapping for
the obsolete schema version, then upgrade the stored mutation
using obtained column mapping and apply an upgraded mutation instead.
In case we don't find a column_mapping we just return an error
from the learn stage.
Tests: unit(dev, debug), dtests(paxos_tests.py:TestPaxos.schema_mismatch_*_test)
Signed-off-by: Pavel Solodovnikov <pa.solodovnikov@scylladb.com>
Try to look up and use schema from the local schema_registry
in case when we have a schema mismatch between the most recent schema
version and the one that is stored inside the `frozen_mutation` for
the accepted proposal.
When such situation happens the stored `frozen_mutation` is able
to be applied only if we are lucky enough and column_mapping in
the mutation is "compatible" with the new table schema.
It wouldn't work if, for example, the columns are reordered, or
some columns, which are referenced by an LWT query, are dropped.
With the patch we are able to mitigate these cases as long as the
referenced schema is still present in the node cache (e.g.
it didn't restart/crash or the cache entry is not too old
to be evicted).
Tests: unit(dev, debug), dtest(paxos_tests.schema_mismatch_*_test)
Signed-off-by: Pavel Solodovnikov <pa.solodovnikov@scylladb.com>
Message-Id: <20200827150844.624017-1-pa.solodovnikov@scylladb.com>
Previously system.paxos TTL was set as max(3h, gc_grace_seconds).
Introduce new per-table option named `paxos_grace_seconds` to set
the amount of seconds which are used to TTL data in paxos tables
when using LWT queries against the base table.
Default value is equal to `DEFAULT_GC_GRACE_SECONDS`,
which is 10 days.
This change allows to easily test various issues related to paxos TTL.
Fixes#6284
Tests: unit (dev, debug)
Co-authored-by: Alejo Sanchez <alejo.sanchez@scylladb.com>
Signed-off-by: Alejo Sanchez <alejo.sanchez@scylladb.com>
Signed-off-by: Pavel Solodovnikov <pa.solodovnikov@scylladb.com>
Message-Id: <20200816223935.919081-1-pa.solodovnikov@scylladb.com>
Use the recently added `max_result_size` field of `query::read_command`
to pass the max result size around, including passing it to remote
nodes. This means that the max result size will be sent along each read,
instead of once per connection.
As we want to select the appropriate `max_result_size` based on the type
of the query as well as based on the query class (user or internal) the
previous method won't do anymore. If the remote doesn't fill this
field, the old per-connection value is used.
This patch changes the per table latencies histograms: read, write,
cas_prepare, cas_accept, and cas_learn.
Beside changing the definition type and the insertion method, the API
was changed to support the new metrics.
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
We were not consistent about using '#include "foo.hh"' instead of
'#include <foo.hh>' for scylla's own headers. This patch fixes that
inconsistency and, to enforce it, changes the build to use -iquote
instead of -I to find those headers.
Signed-off-by: Rafael Ávila de Espíndola <espindola@scylladb.com>
Message-Id: <20200608214208.110216-1-espindola@scylladb.com>
In order to add tracing to places where it can be useful,
e.g. materialized view updates and hinted handoff, tracing state
is propagated to all applicable call sites.
Some caller have partition_key_view, but not partition_key, so thy need
to create a temporary and copy just to pass a reference. Change it by
accepting a view.
Rename inherited metrics cas_propose and cas_commit
to cas_accept and cas_learn respectively.
A while ago we made a decision to stick to widely accepted
terms for Paxos rounds: prepare, accept, learn. The rest
of the code is using these terms, so rename the metrics
to avoid confusion/technical debt.
While at it, rename a few internal methods and functions.
Fixes#6169
Message-Id: <20200414213537.129547-1-kostja@scylladb.com>
The following sleep injections are added to paxos_state:
* paxos_state_prepare_timeout (timeouts in paxos_state::prepare)
* paxos_state_accept_timeout (timeouts in paxos_state::accept)
* paxos_state_learn_timeout (timeouts in paxos_state::learn)
Tests: unit ({dev}), unit ({debug})
Signed-off-by: Pavel Solodovnikov <pa.solodovnikov@scylladb.com>
Signed-off-by: Alejo Sanchez <alejo.sanchez@scylladb.com>
Message-Id: <20200403092107.181057-1-alejo.sanchez@scylladb.com>
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>
and replace all calls to dht::global_partitioner().get_token
dht::get_token is better because it takes schema and uses it
to obtain partitioner instead of using a global partitioner.
Signed-off-by: Piotr Jastrzebski <piotr@scylladb.com>
This will prevent contention in case of parallel updates of the same row
by the same coordinator. The patch does it by introducing a new per key
lock map and taking it before running PAXOS protocol (either for write
of for read).
Message-Id: <20200117101228.GA14816@scylladb.com>
Merged patch series from Gleb Natapov:
"LWT is much more efficient if a request is processed on a shard that owns
a token for the request. This is because otherwise the processing will
bounce to an owning shard multiple times. The patch proposes a way to
move request to correct shard before running lwt. It works by returning
an error from lwt code if a shard is incorrect one specifying the shard
the request should be moved to. The error is processed by the transport
code that jumps to a correct shard and re-process incoming message there.
The nicer way to achieve the same would be to jump to a right shard
inside of the storage_proxy::cas(), but unfortunately with current
implementation of the modification statements they are unusable by
a shard different from where it was created, so the jump should happen
before a modification statement for an cas() is created. When we fix our
cql code to be more cross-shard friendly this can be reworked to do the
jump in the storage_proxy."
Gleb Natapov (4):
transport: change make_result to takes a reference to cql result
instead of shared_ptr
storage_service: move start_native_transport into a thread
lwt: Process lwt request on a owning shard
lwt: drop invoke_on in paxos_state prepare and accept
auth/service.hh | 5 +-
message/messaging_service.hh | 2 +-
service/client_state.hh | 30 +++-
service/paxos/paxos_state.hh | 10 +-
service/query_state.hh | 6 +
service/storage_proxy.hh | 2 +
transport/messages/result_message.hh | 20 +++
transport/messages/result_message_base.hh | 4 +
transport/request.hh | 4 +
transport/server.hh | 25 ++-
cql3/statements/batch_statement.cc | 6 +
cql3/statements/modification_statement.cc | 6 +
cql3/statements/select_statement.cc | 8 +
message/messaging_service.cc | 2 +-
service/paxos/paxos_state.cc | 48 ++---
service/storage_proxy.cc | 47 ++++-
service/storage_service.cc | 120 +++++++------
test/boost/cql_query_test.cc | 1 +
thrift/handler.cc | 3 +
transport/messages/result_message.cc | 5 +
transport/server.cc | 203 ++++++++++++++++------
21 files changed, 377 insertions(+), 180 deletions(-)
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.
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.
This patch adds the following per table stats:
cas_prepare_latency
cas_propose_latency
cas_commit_latency
They are equivalent to CasPropose, CasPrepare, CasCommit metrics exposed
by Cassandra.
This patch adds all functionality needed for Paxos protocol. The
implementation does not strictly adhere to Paxos paper since the original
paper allows setting a value only once, while for LWT we need to be able
to make another Paxos round after "learn" phase completes, which requires
things like repair to be introduced.
Paxos protocol relies on replicas having a state that persists over
crashes/restarts. This patch defines such state and stores it in the
database itself in the paxos table to make it persistent.
The stored state is:
in_progress_ballot - promised ballot
proposal - accepted value
proposal_ballot - the ballot of the accepted value
most_recent_commit - most recently learned value
most_recent_commit_at - the ballot of the most recently learned value
This patch add two data structures that will be used by paxos. First
one is "proposal" which contains a ballot and a mutation representing
a value paxos protocol is trying to set. Second one is
"prepare_response" which is a value returned by paxos prepare stage.
It contains currently accepted value (if any) and most recently
learned value (again if any). The later is used to "repair" replicas
that missed previous "learn" message.
