Commit 8c4b5e4283 introduced an optimization which only
calculates max purgeable timestamp when a tombstone satisfy the
grace period.
Commit 'repair: Get rid of the gc_grace_seconds' inverted the order,
probably under the assumption that getting grace period can be
more expensive than calculating max purgeable, as repair-mode GC
will look up into history data in order to calculate gc_before.
This caused a significant regression on tombstone heavy compactions,
where most of tombstones are still newer than grace period.
A compaction which used to take 5s, now takes 35s. 7x slower.
The reason is simple, now calculation of max purgeable happens
for every single tombstone (once for each key), even the ones that
cannot be GC'ed yet. And each calculation has to iterate through
(i.e. check the bloom filter of) every single sstable that doesn't
participate in compaction.
Flame graph makes it very clear that bloom filter is a heavy path
without the optimization:
45.64% 45.64% sstable_compact sstable_compaction_test_g
[.] utils::filter::bloom_filter::is_present
With its resurrection, the problem is gone.
This scenario can easily happen, e.g. after a deletion burst, and
tombstones becoming only GC'able after they reach upper tiers in
the LSM tree.
Before this patch, a compaction can be estimated to have this # of
filter checks:
(# of keys containing *any* tombstone) * (# of uncompacting sstable
runs[1])
[1] It's # of *runs*, as each key tend to overlap with only one
fragment of each run.
After this patch, the estimation becomes:
(# of keys containing a GC'able tombstone) * (# of uncompacting
runs).
With repair mode for tombstone GC, the assumption, that retrieval
of gc_before is more expensive than calculating max purgeable,
is kept. We can revisit it later. But the default mode, which
is the "timeout" (i.e. gc_grace_seconds) one, we still benefit
from the optimization of deferring the calculation until
needed.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Closes#13908
The immediate mode is similar to timeout mode with gc_grace_seconds
zero. Thus, the gc_before returned should be the query_time instead of
gc_clock::time_point::max in immediate mode.
Setting gc_before to gc_clock::time_point::max, a row could be dropped
by compaction even if the ttl is not expired yet.
The following procedure reproduces the issue:
- Start 2 nodes
- Insert data
```
CREATE KEYSPACE ks2a WITH REPLICATION = { 'class' : 'SimpleStrategy',
'replication_factor' : 2 };
CREATE TABLE ks2a.tb (pk int, ck int, c0 text, c1 text, c2 text, PRIMARY
KEY(pk, ck)) WITH tombstone_gc = {'mode': 'immediate'};
INSERT into ks2a.tb (pk,ck, c0, c1, c2) values (10 ,1, 'x', 'y', 'z')
USING TTL 1000000;
INSERT into ks2a.tb (pk,ck, c0, c1, c2) values (20 ,1, 'x', 'y', 'z')
USING TTL 1000000;
INSERT into ks2a.tb (pk,ck, c0, c1, c2) values (30 ,1, 'x', 'y', 'z')
USING TTL 1000000;
```
- Run nodetool flush and nodetool compact
- Compaction drops all data
```
~128 total partitions merged to 0.
```
Fixes#13572Closes#13800
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
now that C++20 is able to generate the default-generated comparing
operators for us. there is no need to define them manually. and,
`std::rel_ops::*` are deprecated in C++20.
also, use `foo <=> bar` instead of `tri_compare(foo, bar)` for better
readability.
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
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.
The only reason why it's there (right next to compaction_fwd.hh) is
because the database::table_truncate_state subclass needs the definition
of compaction_manager::compaction_reenabler subclass.
However, the former sub is not used outside of database.cc and can be
defined in .cc. Keeping it outside of the header allows dropping the
compaction_manager.hh from database.hh thus greatly reducing its fanout
over the code (from ~180 indirect inclusions down to ~20).
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
Closes#13622
Schema related files are moved there. This excludes schema files that
also interact with mutations, because the mutation module depends on
the schema. Those files will have to go into a separate module.
Closes#12858
Move the thread-local instances of the
per-table repair history maps into compaction_manager.
Fixes#11208
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
and use it to access the repair history maps.
At this introductory patch, we use default-constructed
tombstone_gc_state to access the thread-local maps
temporarily and those use sites will be replaced
in following patches that will gradually pass
the tombstone_gc_state down from the compaction_manager
to where it's used.
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
Define table_id as a distinct utils::tagged_uuid modeled after raft
tagged_id, so it can be differentiated from other uuid-class types,
in particular from table_schema_version.
Fixes#11207
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
Each feature has a private variable and a public accessor. Since the
accessor effectively makes the variable public, avoid the intermediary
and make the variable public directly.
To ease mechanical translation, the variable name is chosen as
the function name (without the cluster_supports_ prefix).
References throughout the codebase are adjusted.
It is typical in switch statements to select on an enum type and
rely on the compliler to complain if an enum value was missed. But
gcc isn't satisified since the enum could have a value outside the
declared list. Call abort() in this impossible situation to pacify
it.
But only on the surface, the only internal function needing the database
(`needs_repair_before_gc()`) still gets a real database because the
replication factor cannot be obtained from the data dictionary
currently. Although this might not look like an improvement, it is
enough to avoid a `real_database()` call for tables that don't have
tombstone gc mode set to repair.
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.
The gc_grace_seconds is a very fragile and broken design inherited from
Cassandra. Deleted data can be resurrected if cluster wide repair is not
performed within gc_grace_seconds. This design pushes the job of making
the database consistency to the user. In practice, it is very hard to
guarantee repair is performed within gc_grace_seconds all the time. For
example, repair workload has the lowest priority in the system which can
be slowed down by the higher priority workload, so that there is no
guarantee when a repair can finish. A gc_grace_seconds value that is
used to work might not work after data volume grows in a cluster. Users
might want to avoid running repair during a specific period where
latency is the top priority for their business.
To solve this problem, an automatic mechanism to protect data
resurrection is proposed and implemented. The main idea is to remove the
tombstone only after the range that covers the tombstone is repaired.
In this patch, a new table option tombstone_gc is added. The option is
used to configure tombstone gc mode. For example:
1) GC a tombstone after gc_grace_seconds
cqlsh> ALTER TABLE ks.cf WITH tombstone_gc = {'mode':'timeout'} ;
This is the default mode. If no tombstone_gc option is specified by the
user. The old gc_grace_seconds based gc will be used.
2) Never GC a tombstone
cqlsh> ALTER TABLE ks.cf WITH tombstone_gc = {'mode':'disabled'};
3) GC a tombstone immediately
cqlsh> ALTER TABLE ks.cf WITH tombstone_gc = {'mode':'immediate'};
4) GC a tombstone after repair
cqlsh> ALTER TABLE ks.cf WITH tombstone_gc = {'mode':'repair'};
In addition to the 'mode' option, another option 'propagation_delay_in_seconds'
is added. It defines the max time a write could possibly delay before it
eventually arrives at a node.
A new gossip feature TOMBSTONE_GC_OPTIONS is added. The new tombstone_gc
option can only be used after the whole cluster supports the new
feature. A mixed cluster works with no problem.
Tests: compaction_test.py, ninja test
Fixes#3560
[avi: resolve conflicts vs data_dictionary]
