Fixes root cause of a cascade of failures in replication code:
- Replica handling of deleting an edge is now corrected. Now tolerant of multiple edges of the same relationship type.
- Improved robustness: correct exception handling around failed stream of current WAL file. This now means a REPLICA failure will no longer prevent transactions on MAIN from performing WAL writes.
- Slightly better diagnostic messages, not user friendly but helps get developer to correct root cause quicker.
- Proactively remove vertex+edges during Abort rather than defer to GC to do that work, this included fixing constraints and indexes to be safe.
Co-authored-by: Andreja Tonev <andreja.tonev@memgraph.io>
Single (instance level) connection to a replica (messages from all databases get multiplexed through it)
ReplicationClient split in two: ReplicationClient and ReplicationStorageClient
New ReplicationClient, moved under replication, handles the raw connection, owned by MainRoleData
ReplicationStorageClient handles the storage <-> replica state machine and holds to a stream
Removed epoch and storage from *Clients
rpc::Stream proactively aborts on error and sets itself to a defunct state
Removed HandleRpcFailure, instead we simply log the error and let the FrequentCheck handle re-connection
replica_state is now a synced variable
ReplicaStorageClient state machine bugfixes
Single FrequentCheck that goes through DBMS
Moved ReplicationState under DbmsHandler
Moved some replication startup logic under the DbmsHandler's constructor
Removed InMemoryReplicationClient
CreateReplicationClient has been removed from Storage
Simplified GetRecoverySteps and made safer
---------
Co-authored-by: Gareth Lloyd <gareth.lloyd@memgraph.io>
There might be a performance impect of updating the metadata store on
bulk operations. Hence this flag which is disabling the collection by
default. If the queries to obtain the information are called with this
flag disabled, the database will throw an exception.
The objects stored_node_labels_ and stored_edge_types_ can be accesses
through separate threads but it was not safe to do so. This commit
replaces the standard containers with threadsafe ones.
Before the functions that are retrieving the data from the metadata
holding datastructures were returning a std::string, and that was
propagated outward all the way through. To keep this functions
consistent with the rest of the storage/dbaccessor functions the LabelId
and EdgeTypeId will be propagated respectively and the conversion into
string will only happen at the interpreter level.
In order to get the required metadata in constant time we need to keep
track of the node labels and edge types that were ever present in the
database. This is done by the two axuiliary datastructures that are
present in the storage instances. The ability to get this metadata is
propagated to the DBAccessor class, which the query modules can interact
with.
* Interpreter transaction ID decoupled from storage transaction ID
* Transactional scope for indices, statistics and constraints
* Storage::Accessor now has 2 modes (unique and shared)
* Introduced ResourceLock to fix pthread mutex problems
* Split InfoQuery in two: non-transactional SystemInfoQuery and transactional DatabaseInfoQuery
* Replicable and durable statistics
* Bumped WAL/Snapshot versions
* Initial implementation of the Lamport clock
---------
Co-authored-by: Andreja Tonev <andreja.tonev@memgraph.io>
Unique/global InterpreterContext that is Storage agnostic (has a reference to the DbmsHandler instead)
* InterpreterContext is no longer the owner of Storage
* New Database structure that handles Storage, Triggers, Streams
* Renamed SessinContextHandler to DbmsHandler and simplified the multi-tenant logic
* Added Gatekeeper and updated handlers to use it
---------
Co-authored-by: Gareth Lloyd <gareth.lloyd@memgraph.io>
