Previously the generation committed in `commit_cdc_generation` state
would be published by the coordinator in `write_both_read_old` state.
This logic assumed that we only create new CDC generations during node
bootstrap.
We'll allow committing new generations without bootstrap (without any
node transitions in fact), so we need this separate state.
After publishing the generation, we check whether there is a
transitioning node; if so, we'll enter `write_both_read_old` as next
state, otherwise we'll make the topology non-transitioning.
This function broadcasts a command to cluster members. It takes a
`node_to_work_on`. We'll need a version which works in situations where
there is no 'node to work on'.
This calls `raft_group0_client::start_operation` and checks if the term
is different from the term that the coordinator was initially created
with; if so, we must no longer continue coordinating the topology.
There was one direct call to `raft_group0_client::start_operation`
without a term check, replace it with the introduced function.
The existing `topology_mutation_builder` took a `raft::server_id` in its
constructor and immediately created a clustering row in the
`system.topology` mutation that it was building for the given node.
This does not allow building mutations which only affect the static
columns.
Split the class into two:
- `topology_mutation_builder` doesn't take `raft::server_id` in its
constructor and contains only the methods that are used to set static
columns. It also has a `with_node` method taking a `raft::server_id`
which returns a `topology_node_mutation_builder&`.
- `topology_node_mutation_builder` creates the clustering row and allows
seting its columns.
We'll use `topology_mutation_builder` when we only want to transition
the cluster-global topology state, without affecting any specific nodes'
states.
`topology` currently contains the `requests` map, which is suitable for
node-specific requests such as "this node wants to join" or "this node
must be removed". But for requests for operations that affect the
cluster as a whole, a separate request type and field is more
appropriate. Introduce one.
The enum currently contains the option `new_cdc_generation` for requests
to create a new CDC generation in the cluster. We will implement the
whole procedure in later commits.
- make it a static column in `system.topology`
- move it from node-specific `ring_slice` to cluster-global `topology`
We will use it in scenarios where no node is transitioning.
Also make it `std::optional` in topology for consistency with other
fields (previously, the 'no value' state for this field was represented
using default-constructed `utils::UUID`).
For some reason Scylla crashes on `aarch64` in release mode when calling
`fmt::format` in `raft_removenode` and `raft_decommission`. E.g. on this
line:
```
group0_command g0_cmd = _group0->client().prepare_command(std::move(change), guard, fmt::format("decomission: request decomission for {}", raft_server.id()));
```
I found this in our configure.py:
```
def get_clang_inline_threshold():
if args.clang_inline_threshold != -1:
return args.clang_inline_threshold
elif platform.machine() == 'aarch64':
# we see miscompiles with 1200 and above with format("{}", uuid)
# also coroutine miscompiles with 600
return 300
else:
return 2500
```
but reducing it to `0` didn't help.
I managed to get the following backtrace (with inline threshold 0):
```
void boost::intrusive::list_impl<boost::intrusive::mhtraits<seastar::thread_context, boost::intrusive::list_member_hook<>, &seastar::thread_context::_all_link>, unsigned long, false, void>::clear_and_dispose<boost::intrusive::detail::null_disposer>(boost::intrusive::detail::null_disposer) at /usr/include/boost/intrusive/list.hpp:751
(inlined by) boost::intrusive::list_impl<boost::intrusive::mhtraits<seastar::thread_context, boost::intrusive::list_member_hook<>, &seastar::thread_context::_all_link>, unsigned long, false, void>::clear() at /usr/include/boost/intrusive/list.hpp:728
(inlined by) ~list_impl at /usr/include/boost/intrusive/list.hpp:255
void fmt::v9::detail::buffer<wchar_t>::append<wchar_t>(wchar_t const*, wchar_t const*) at ??:?
void fmt::v9::detail::vformat_to<char>(fmt::v9::detail::buffer<char>&, fmt::v9::basic_string_view<char>, fmt::v9::basic_format_args<fmt::v9::basic_format_context<std::conditional<std::is_same<fmt::v9::type_identity<char>::type, char>::value, fmt::v9::appender, std::back_insert_iterator<fmt::v9::detail::buffer<fmt::v9::type_identity<char>::type> > >::type, fmt::v9::type_identity<char>::type> >, fmt::v9::detail::locale_ref) at ??:?
fmt::v9::vformat[abi:cxx11](fmt::v9::basic_string_view<char>, fmt::v9::basic_format_args<fmt::v9::basic_format_context<fmt::v9::appender, char> >) at ??:?
std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > fmt::v9::format<utils::tagged_uuid<raft::server_id_tag>&>(fmt::v9::basic_format_string<char, fmt::v9::type_identity<utils::tagged_uuid<raft::server_id_tag>&>::type>, utils::tagged_uuid<raft::server_id_tag>&) at /usr/include/fmt/core.h:3206
(inlined by) service::storage_service::raft_removenode(utils::tagged_uuid<locator::host_id_tag>) at ./service/storage_service.cc:3572
```
Maybe it's a bug in `fmt` library?
In any case replacing the call with `::format` (i.e. `seastar::format`
from seastar/core/print.hh) helps.
Do it for the entire file for consistency (and avoiding this bug).
Also, for the future, replace `format` calls with `::format` - now it's
the same thing, but the latter won't clash with `std::format` once we
switch to libstdc++13.
Fixes#13707Closes#13711
storage_service uses raft_group0 but the during shutdown the later is
destroyed before the former is stopped. This series move raft_group0
destruction to be after storage_service is stopped already. For the
move to work some existing dependencies of raft_group0 are dropped
since they do not really needed during the object creation.
Fixes#13522
raft_group0 does not really depends on cdc::generation_service, it needs
it only transiently, so pass it to appropriate methods of raft_group0
instead of during its creation.
We change the meaning and name of `replication_state`: previously it was meant
to describe the "state of tokens" of a specific node; now it describes the
topology as a whole - the current step in the 'topology saga'. It was moved
from `ring_slice` into `topology`, renamed into `transition_state`, and the
topology coordinator code was modified to switch on it first instead of node
state - because there may be no single transitioning node, but the topology
itself may be transitioning.
This PR was extracted from #13683, it contains only the part which refactors
the infrastructure to prepare for non-node specific topology transitions.
Closes#13690
* github.com:scylladb/scylladb:
raft topology: rename `update_replica_state` -> `update_topology_state`
raft topology: remove `transition_state::normal`
raft topology: switch on `transition_state` first
raft topology: `handle_ring_transition`: rename `res` to `exec_command_res`
raft topology: parse replaced node in `exec_global_command`
raft topology: extract `cleanup_group0_config_if_needed` from `get_node_to_work_on`
storage_service: extract raft topology coordinator fiber to separate class
raft topology: rename `replication_state` to `transition_state`
raft topology: make `replication_state` a topology-global state
Consider
- n1, n2, n3
- n3 is down
- n4 replaces n3 with the same ip address 127.0.0.3
- Inside the storage_service::handle_state_normal callback for 127.0.0.3 on n1/n2
```
auto host_id = _gossiper.get_host_id(endpoint);
auto existing = tmptr->get_endpoint_for_host_id(host_id);
```
host_id = new host id
existing = empty
As a result, del_replacing_endpoint() will not be called.
This means 127.0.0.3 will not be removed as a pending node on n1 and n2 when
replacing is done. This is wrong.
This is a regression since commit 9942c60d93
(storage_service: do not inherit the host_id of a replaced a node), where
replacing node uses a new host id than the node to be replaced.
To fix, call del_replacing_endpoint() when a node becomes NORMAL and existing
is empty.
Before:
n1:
storage_service - replace[cd1f187a-0eee-4b04-91a9-905ecc499cfc]: Added replacing_node=127.0.0.3 to replace existing_node=127.0.0.3, coordinator=127.0.0.3
token_metadata - Added node 127.0.0.3 as pending replacing endpoint which replaces existing node 127.0.0.3
storage_service - replace[cd1f187a-0eee-4b04-91a9-905ecc499cfc]: Marked ops done from coordinator=127.0.0.3
storage_service - Node 127.0.0.3 state jump to normal
storage_service - Set host_id=6f9ba4e8-9457-4c76-8e2a-e2be257fe123 to be owned by node=127.0.0.3
After:
n1:
storage_service - replace[28191ea6-d43b-3168-ab01-c7e7736021aa]: Added replacing_node=127.0.0.3 to replace existing_node=127.0.0.3, coordinator=127.0.0.3
token_metadata - Added node 127.0.0.3 as pending replacing endpoint which replaces existing node 127.0.0.3
storage_service - replace[28191ea6-d43b-3168-ab01-c7e7736021aa]: Marked ops done from coordinator=127.0.0.3
storage_service - Node 127.0.0.3 state jump to normal
token_metadata - Removed node 127.0.0.3 as pending replacing endpoint which replaces existing node 127.0.0.3
storage_service - Set host_id=72219180-e3d1-4752-b644-5c896e4c2fed to be owned by node=127.0.0.3
Tests: https://github.com/scylladb/scylla-dtest/pull/3126Closes#13677
The new name is more generic and appropriate for topology transitions
which don't affect any specific replica but the entire cluster as a
whole (which we'll introduce later).
Also take `guard` directly instead of `node_to_work_on` in this more
generic function. Since we want `node_to_work_on` to die when we steal
its guard, introduce `take_guard` which takes ownership of the object
and returns the guard.
Previously the code assumed that there was always a 'node to work on' (a
node which wants to change its state) or there was no work to do at all.
It would find such a node, switch on its state (e.g. check if it's
bootstrapping), and in some states switch on the topology
`transition_state` (e.g. check if it's `write_both_read_old`).
We want to introduce transitions that are not node-specific and can work
even when all nodes are 'normal' (so there's no 'node to work on'). As a
first step, we refactor the code so it switches on `transition_state`
first. In some of these states, like `write_both_read_old`, there must
be a 'node to work on' for the state to make sense; but later in some
states it will be optional (such as `commit_cdc_generation`).
The lambdas defined inside the fiber are now methods of this class.
Currently `handle_node_transition` is calling `handle_ring_transition`,
in a later commit we will reverse this: `handle_ring_transition` will
call `handle_node_transition`. We won't have to shuffle the functions
around because they are members of the same class, making the change
easier to review. In general, the code will be easier to maintain in
this new form (no need to deal with so many lambda captures etc.)
Also break up some lines which exceeded the 120 character limit (as per
Seastar coding guidelines).
The new name is more generic - it describes the current step of a
'topology saga` (a sequence of steps used to implement a larger topology
operation such as bootstrap).
Previously it was part of `ring_slice`, belonging to a specific node.
This commit moves it into `topology`, making it a cluster-global
property.
The `replication_state` column in `system.topology` is now `static`.
This will allow us to easily introduce topology transition states that
do not refer to any specific node. `commit_cdc_generation` will be such
a state, allowing us to commit a new CDC generation even though all
nodes are normal (none are transitioning). One could argue that the
other states are conceptually already cluster-global: for example,
`write_both_read_new` doesn't affect only the tokens of a bootstrapping
(or decommissioning etc.) node; it affects replica sets of other tokens
as well (with RFs greater than 1).
This PR introduces an experimental feature called "tablets". Tablets are
a way to distribute data in the cluster, which is an alternative to the
current vnode-based replication. Vnode-based replication strategy tries
to evenly distribute the global token space shared by all tables among
nodes and shards. With tablets, the aim is to start from a different
side. Divide resources of replica-shard into tablets, with a goal of
having a fixed target tablet size, and then assign those tablets to
serve fragments of tables (also called tablets). This will allow us to
balance the load in a more flexible manner, by moving individual tablets
around. Also, unlike with vnode ranges, tablet replicas live on a
particular shard on a given node, which will allow us to bind raft
groups to tablets. Those goals are not yet achieved with this PR, but it
lays the ground for this.
Things achieved in this PR:
- You can start a cluster and create a keyspace whose tables will use
tablet-based replication. This is done by setting `initial_tablets`
option:
```
CREATE KEYSPACE test WITH replication = {'class': 'NetworkTopologyStrategy',
'replication_factor': 3,
'initial_tablets': 8};
```
All tables created in such a keyspace will be tablet-based.
Tablet-based replication is a trait, not a separate replication
strategy. Tablets don't change the spirit of replication strategy, it
just alters the way in which data ownership is managed. In theory, we
could use it for other strategies as well like
EverywhereReplicationStrategy. Currently, only NetworkTopologyStrategy
is augmented to support tablets.
- You can create and drop tablet-based tables (no DDL language changes)
- DML / DQL work with tablet-based tables
Replicas for tablet-based tables are chosen from tablet metadata
instead of token metadata
Things which are not yet implemented:
- handling of views, indexes, CDC created on tablet-based tables
- sharding is done using the old method, it ignores the shard allocated in tablet metadata
- node operations (topology changes, repair, rebuild) are not handling tablet-based tables
- not integrated with compaction groups
- tablet allocator piggy-backs on tokens to choose replicas.
Eventually we want to allocate based on current load, not statically
Closes#13387
* github.com:scylladb/scylladb:
test: topology: Introduce test_tablets.py
raft: Introduce 'raft_server_force_snapshot' error injection
locator: network_topology_strategy: Support tablet replication
service: Introduce tablet_allocator
locator: Introduce tablet_aware_replication_strategy
locator: Extract maybe_remove_node_being_replaced()
dht: token_metadata: Introduce get_my_id()
migration_manager: Send tablet metadata as part of schema pull
storage_service: Load tablet metadata when reloading topology state
storage_service: Load tablet metadata on boot and from group0 changes
db, migration_manager: Notify about tablet metadata changes via migration_listener::on_update_tablet_metadata()
migration_notifier: Introduce before_drop_keyspace()
migration_manager: Make prepare_keyspace_drop_announcement() return a future<>
test: perf: Introduce perf-tablets
test: Introduce tablets_test
test: lib: Do not override table id in create_table()
utils, tablets: Introduce external_memory_usage()
db: tablets: Add printers
db: tablets: Add persistence layer
dht: Use last_token_of_compaction_group() in split_token_range_msb()
locator: Introduce tablet_metadata
dht: Introduce first_token()
dht: Introduce next_token()
storage_proxy: Improve trace-level logging
locator: token_metadata: Fix confusing comment on ring_range()
dht, storage_proxy: Abstract token space splitting
Revert "query_ranges_to_vnodes_generator: fix for exclusive boundaries"
db: Exclude keyspace with per-table replication in get_non_local_strategy_keyspaces_erms()
db: Introduce get_non_local_vnode_based_strategy_keyspaces()
service: storage_proxy: Avoid copying keyspace name in write handler
locator: Introduce per-table replication strategy
treewide: Use replication_strategy_ptr as a shorter name for abstract_replication_strategy::ptr_type
locator: Introduce effective_replication_map
locator: Rename effective_replication_map to vnode_effective_replication_map
locator: effective_replication_map: Abstract get_pending_endpoints()
db: Propagate feature_service to abstract_replication_strategy::validate_options()
db: config: Introduce experimental "TABLETS" feature
db: Log replication strategy for debugging purposes
db: Log full exception on error in do_parse_schema_tables()
db: keyspace: Remove non-const replication strategy getter
config: Reformat
in C++20, compiler generate operator!=() if the corresponding
operator==() is already defined, the language now understands
that the comparison is symmetric in the new standard.
fortunately, our operator!=() is always equivalent to
`! operator==()`, this matches the behavior of the default
generated operator!=(). so, in this change, all `operator!=`
are removed.
in addition to the defaulted operator!=, C++20 also brings to us
the defaulted operator==() -- it is able to generated the
operator==() if the member-wise lexicographical comparison.
under some circumstances, this is exactly what we need. so,
in this change, if the operator==() is also implemented as
a lexicographical comparison of all memeber variables of the
class/struct in question, it is implemented using the default
generated one by removing its body and mark the function as
`default`. moreover, if the class happen to have other comparison
operators which are implemented using lexicographical comparison,
the default generated `operator<=>` is used in place of
the defaulted `operator==`.
sometimes, we fail to mark the operator== with the `const`
specifier, in this change, to fulfil the need of C++ standard,
and to be more correct, the `const` specifier is added.
also, to generate the defaulted operator==, the operand should
be `const class_name&`, but it is not always the case, in the
class of `version`, we use `version` as the parameter type, to
fulfill the need of the C++ standard, the parameter type is
changed to `const version&` instead. this does not change
the semantic of the comparison operator. and is a more idiomatic
way to pass non-trivial struct as function parameters.
please note, because in C++20, both operator= and operator<=> are
symmetric, some of the operators in `multiprecision` are removed.
they are the symmetric form of the another variant. if they were
not removed, compiler would, for instance, find ambiguous
overloaded operator '=='.
this change is a cleanup to modernize the code base with C++20
features.
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
Closes#13687
Fix two issues with the replace operation introduced by recent PRs.
Add a test which performs a sequence of basic topology operations (bootstrap,
decommission, removenode, replace) in a new suite that enables the `raft`
experimental feature (so that the new topology change coordinator code is used).
Fixes: #13651Closes#13655
* github.com:scylladb/scylladb:
test: new suite for testing raft-based topology
test: remove topology_custom/test_custom.py
raft topology: don't require new CDC generation UUID to always be present
raft topology: include shard_count/ignore_msb during replace
this is a part of a series to migrating from `operator<<(ostream&, ..)`
based formatting to fmtlib based formatting. the goal here is to enable
fmtlib to print `storage_service::mode` without the help of `operator<<`.
the corresponding `operator<<()` for `storage_service::mode` is removed
in this change, as all its callers are now using fmtlib for formatting
now.
Refs #13245
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
Closes#13640
raft_group0 does not really depends on migration_manager, it needs it only
transiently, so pass it to appropriate methods of raft_group0 instead
of during its creation.
raft_group0 does not really depends on query_processor, it needs it only
transiently, so pass it to appropriate methods of raft_group0 instead
of during its creation.
raft_group0 does not really depends on storage_service, it needs it only
transiently, so pass it to appropriate methods of raft_group0 instead
of during its creation.
During node replace we don't introduce a new CDC generation, only during
regular bootstrap. Instead of checking that `new_cdc_generation_uuid`
must be present whenever there's a topology transition, only check it
when we're in `commit_cdc_generation` state.
Currently, responsible for injecting mutations of system.tablets to
schema changes.
Note that not all migrations are handled currently. Dependant view or
cdc table drops are not handled.
This change puts the reloading into topology_state_load(), which is a
function which reloads token_metadata from system.topology (the new
raft-based topology management). It clears the metadata, so needs to
reload tablet map too. In the future, tablet metadata could change as
part of topology transaction too, so we reload rather than preserve.
Currently, scans are splitting partition ranges around tokens. This
will have to change with tablets, where we should split at tablet
boundaries.
This patch introduces token_range_splitter which abstracts this
task. It is provided by effective_replication_map implementation.
It's meant to be used in places where currently
get_non_local_strategy_keyspaces() is used, but work only with
keyspaces which use vnode-based replication strategy.
Will be used by tablet-based replication strategies, for which
effective replication map is different per table.
Also, this patch adapts existing users of effective replication map to
use the per-table effective replication map.
For simplicity, every table has an effective replication map, even if
the erm is per keyspace. This way the client code can be uniform and
doesn't have to check whether replication strategy is per table.
Not all users of per-keyspace get_effective_replication_map() are
adapted yet to work per-table. Those algorithms will throw an
exception when invoked on a keyspace which uses per-table replication
strategy.
All users of global proxy are gone (*), proxy can be made fully main/cql_test_env local.
(*) one test case still needs it, but can get it via cql_test_env
Closes#13616
* github.com:scylladb/scylladb:
code: Remove global proxy
schema_change_test: Use proxy from cql_test_env
test: Carry proxy reference on cql_test_env
Derived from utils::tagged_integer, using different tags,
the types are incompatible with each other and require explicit
typecasting to- and from- their value type.
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
Although get_generation_number implementation is
completely generic, it is used exclusively to seed
the gossip generation number.
Following patches will define a strong gms::generation_id
type and this function should return it.
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
and provide accessor functions to get them.
1. So they can't be modified by mistake, as the versioned value is
immutable. A new value must have a higher version.
2. Before making the version a strong gms::version_type.
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
