Nearest dc balancer #206
Nearest dc balancer #206
Conversation
|
…loop if continue from error. Added tests
ydb/src/load_balancer.rs
Outdated
| } | ||
|
|
||
| const NODES_PER_DC: usize = 5; | ||
| const PING_TIMEOUT_SECS: u64 = 60; |
There was a problem hiding this comment.
what about use discovery/ping interval for ping timeout?
There was a problem hiding this comment.
60 sec - interval for TimerDiscovery renewal (there is also no predefined constants)
https://github.com/ydb-platform/ydb-rs-sdk/blob/master/ydb/src/client_builder.rs#L283
|
|
||
| impl NearestDCBalancer { | ||
| fn get_endpoint(&self, service: Service) -> YdbResult<Uri> { | ||
| match self.balancer_state.try_lock() { |
There was a problem hiding this comment.
use lock instead of try_lock: we don`t need fast fail at the path: if code doesn't see endpoint it will return to retrier. The retraier will see custom error and fail transaction.
Better wait until update state finish: it is fast operation.
the method should return a error is the ballancer is empty
| }); | ||
| match Self::find_local_dc(&to_check).await { | ||
| Ok(dc) => { | ||
| info!("found new local dc:{}", dc); |
There was a problem hiding this comment.
it is often message, not intresting when all ok. Better to use debug for the message
| ping_token: CancellationToken, | ||
| waiter: Arc<WaiterImpl>, | ||
| config: BalancerConfig, | ||
| balancer_state: Arc<Mutex<BalancerState>>, |
There was a problem hiding this comment.
what about use RWLock instead of Mutex?
it will allow multiply reads in parallel and block for update state only
| new_nodes: &Vec<NodeInfo>, | ||
| local_dc: String, | ||
| ) { | ||
| info!("adjusting endpoints"); |
There was a problem hiding this comment.
it should be debug or trace log level
| tokio::select! { | ||
| biased; // check timeout first | ||
| _ = interrupt_collector_future.cancelled() =>{ | ||
| Self::join_all(&mut nursery).await; // Children will be cancelled due to tokens chaining, see (*) |
No description provided.