5 Commits (bcdc807dd34d7a4e2e1dcfdffecd869da66e6efe)

Author	SHA1	Message	Date
Christian Haudum	ce81895241	Add shuffle sharding to index gateway (#9710) ... This PR introduces shuffle sharding of the index files across index gateways where the shuffle sharding key is the tenant ID. Instead of having a global replication factor on the index gateways which affects all tenants the same, whether they are small or large tenants, the shuffle sharding allows to specify a shard factor per tenant (on top of the replication factor). This a) reduces the amount of indexes that are downloaded on individual index gateways in case of a high replication factor (which is currently the de-factor standard to deal with high load) and b) allows to scale tenants individually. The shuffle sharding algorithm is part of dskit's ring implementation and is not part of this commit. If you have set a high RF for the index gateways, it is required to reduce the RF and instead set the shard factor for the tenants. Signed-off-by: Christian Haudum <christian.haudum@gmail.com>	3 years ago
Dylan Guedes	440fb7d74c	Loki: Modifies TableManager to use IndexGateway ring (#5972) ... * Add two metrics to the IndexGateway. - Add a new `query_readiness_duration_seconds` metric, that reports query readiness duration of a tablemanager/index gateway instance. We should use it later to report performance against the ring mode - Add a new `usersToBeQueryReadyForTotal` metric, that reports number of users involved in the query readiness operation. We should use it later to correlate number of users with the query readiness duration. * Remove `usersToBeQueryReadyForTotal`. - It will report all users always for now, so it isn't too helpful the way it is. * Rename metric help text to not mislead people. * Log queryReadiness duration. * Fix where log message and duration and triggered. * Add two metrics to the IndexGateway. - Add a new `query_readiness_duration_seconds` metric, that reports query readiness duration of a tablemanager/index gateway instance. We should use it later to report performance against the ring mode - Add a new `usersToBeQueryReadyForTotal` metric, that reports number of users involved in the query readiness operation. We should use it later to correlate number of users with the query readiness duration. * Remove `usersToBeQueryReadyForTotal`. - It will report all users always for now, so it isn't too helpful the way it is. * Rename metric help text to not mislead people. * Log queryReadiness duration. * Fix where log message and duration and triggered. * Use boundaries to skip users in TableManager. - Separate the assigning of indexClient in the IndexGateway to allow initializing the shipper with the IndexGateway ring - Add new TenantBoundariesClient entity, that answers if a given tenantID should be ignored or not - Use the TenantBoundariesClient implemented by the IndexGateway in the downloads TableManager * Add IndexGateway configuration docs. * Add tests for pkg/util/ring TokenFor() and IsAssignedKey() Signed-off-by: JordanRushing <rushing.jordan@gmail.com> * Small whitespace fix in pkg/util/ring_test.go Signed-off-by: JordanRushing <rushing.jordan@gmail.com> * Apply suggestions from code review Rewrite QueryIndex error phrase. Co-authored-by: JordanRushing <rushing.jordan@gmail.com> * Join users list in a single string. - This is necessary since go-kit doesn't support array type. * Tweak queryReadiness log messages. - As suggested by Ed on https://github.com/grafana/loki/pull/5972#discussion_r859734129 and https://github.com/grafana/loki/pull/5972#discussion_r859736072 * Implement new RingManager. - Adds a new entity to indexgateway named RingManager, responsible for managing the ring and the lifecycler used by the indexgateway. The ringManager is guaranteed to be initiatiated before the Shipper and before the IndexGateway. - Remove the readiness logic from the IndexGateway. - Pass the RingManager as the TenantBoundaries implementation of the Shipper * Return non-empty ringmanager for all modes. * Fix lint regarding error not being check. * Fix lint due to wrong import order. * Implement support for client and server mode for the ring manager. * Fix ring manager services registration. * Add option to configure whether or not to log gateway requests. * Check if IndexGateway is enabled instead of active. * Tune RingManager behavior. - Instantiate ring buffers inside IsAssignedKey instead of reusing same buffer to avoid issues with concurrency. - Remove unnecessary details from IndexGateway mode docs. - Return error when wrongly instantiating a RingManager when IndexGateway is in simple mode. * Rewrite `TenantInBoundaries` as a func instead of interface. * Fix lint. - Fix YAML tag - Remove else clause * Use distributor instead of querier in test. - Since the querier needs to wait for the index gateway ring, it isn't suitable for this test anymore. * Add ring mode guard clause. * Return code erased by mistake. * Rename TenantInBoundaries to IndexGatewayOwnsTenant. * Log len of users instead of all of them to avoid log msgs. * Add docstrings to new public functions. * Apply suggestions from code review Document that tenant filtering is only applied during query readiness. Co-authored-by: Sandeep Sukhani <sandeep.d.sukhani@gmail.com> * Modify IsAssignedKey to expect address directly instead of lifecycler. - Also removes HasDedicatedAddress, since it isn't necessary anymore * Log error message when IsInReplicationSet fails. * Modify IsAssignedKey to return true by default. * Remove wrong assigning of Service for clientMode. * Log gateway requests before error checks. * Remove unnecessary endpoint registration. * Fix lint. * Update pkg/util/ring.go Co-authored-by: Sandeep Sukhani <sandeep.d.sukhani@gmail.com> Co-authored-by: JordanRushing <rushing.jordan@gmail.com> Co-authored-by: Sandeep Sukhani <sandeep.d.sukhani@gmail.com>	4 years ago
Susana Ferreira	2d24e2ea64	Migration to dskit/ring (#4641)	4 years ago
Ed Welch	fe8bc9144f	Loki: query scheduler should send shutdown to frontends when ReplicationSet changes (#4614) ... * make sure we send a shutdown message to frontends from the scheduler when the ReplicationSet changes so that inflight queries are canceled and retried in the frontends. * use Write so we find an instance incase one is unhealthy	4 years ago
Sandeep Sukhani	2a8a03795a	Improvements in live tailing of logs (#541) ... * Improvements in live tailing of logs Instead of polling for new logs, a grpc stream is opened between ingester and querier to get live logs Querier reconnects to disconnected or newly added ingesters * Added more comments to code for live log tailing * Some code refactoring in live log tailing * handling delayfor in logcli, max delayfor to be 5 seconds * some changes in tail response in live tailing * Fixed issue with stopping ingesters gracefully when live tailing is being used * Added tests for tailer in querier * Live tailing made logql compatible, some code refactoring suggested in PR	7 years ago