Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Upstream contributions from Union.ai (#5769)
* Overlap create execution blob store reads/writes This change modifies launch paths stemming from `launchExecutionAndPrepareModel` to overlap blob store write and read calls, which dominate end-to-end latency (as seen in the traces below). Signed-off-by: Andrew Dye <[email protected]> * Overlap FutureFileReader blob store writes/reads This change updates `FutureFileReader.Cache` and `FutureFileReader.RetrieveCache` to use overlapped write and reads, respectively, to reduce end-to-end latency. The read path is a common operation on each iteration of the propeller `Handle` loop for dynamic nodes. Signed-off-by: Andrew Dye <[email protected]> * Fix async notifications tests I didn't chase down why assumptions changed here and why these tests broke, but fixing them with more explicit checks. Signed-off-by: Andrew Dye <[email protected]> * Overlap fetching input and output data This change updates `GetExecutionData`, `GetNodeExecutionData`, and `GetTaskExecutionData` to use overlapped reads when fetching input and output data. Signed-off-by: Andrew Dye <[email protected]> * Add configuration for launchplan cache resync duration Currently, the launchplan cache resync duration uses the DownstreamEval duration configuration which is also used for the sync period on the k8s client. This means if we want to configure a more aggressive launchplan cache resync, we would also incur overhead in syncing all k8s resources (ex. Pods from `PodPlugin`). By adding a separate configuration value we can update these independently. Signed-off-by: Andrew Dye <[email protected]> * Enqueue owner on launchplan terminal state This PR enqueues the owner workflow for evaluation when the launchplan auto refresh cache detects a launchplan in a terminal state. Signed-off-by: Andrew Dye <[email protected]> * Add client-go metrics Register a few metric callbacks with the client-go metrics interface so that we can monitor request latencies and rate limiting of kubeclient. ``` ❯ curl http://localhost:10254/metrics | rg k8s_client k8s_client_rate_limiter_latency_bucket{verb="GET",le="0.005"} 84 k8s_client_rate_limiter_latency_bucket{verb="GET",le="0.01"} 87 k8s_client_rate_limiter_latency_bucket{verb="GET",le="0.025"} 89 k8s_client_rate_limiter_latency_bucket{verb="GET",le="0.05"} 99 k8s_client_rate_limiter_latency_bucket{verb="GET",le="0.1"} 114 k8s_client_rate_limiter_latency_bucket{verb="GET",le="0.25"} 117 k8s_client_rate_limiter_latency_bucket{verb="GET",le="0.5"} 117 k8s_client_rate_limiter_latency_bucket{verb="GET",le="1"} 117 k8s_client_rate_limiter_latency_bucket{verb="GET",le="2.5"} 117 k8s_client_rate_limiter_latency_bucket{verb="GET",le="5"} 117 k8s_client_rate_limiter_latency_bucket{verb="GET",le="10"} 117 k8s_client_rate_limiter_latency_bucket{verb="GET",le="+Inf"} 117 k8s_client_rate_limiter_latency_sum{verb="GET"} 1.9358371670000003 k8s_client_rate_limiter_latency_count{verb="GET"} 117 k8s_client_rate_limiter_latency_bucket{verb="POST",le="0.005"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="0.01"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="0.025"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="0.05"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="0.1"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="0.25"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="0.5"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="1"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="2.5"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="5"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="10"} 6 k8s_client_rate_limiter_latency_bucket{verb="POST",le="+Inf"} 6 k8s_client_rate_limiter_latency_sum{verb="POST"} 1.0542e-05 k8s_client_rate_limiter_latency_count{verb="POST"} 6 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="0.005"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="0.01"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="0.025"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="0.05"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="0.1"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="0.25"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="0.5"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="1"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="2.5"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="5"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="10"} 1 k8s_client_rate_limiter_latency_bucket{verb="PUT",le="+Inf"} 1 k8s_client_rate_limiter_latency_sum{verb="PUT"} 5e-07 k8s_client_rate_limiter_latency_count{verb="PUT"} 1 k8s_client_request_latency_bucket{verb="GET",le="0.005"} 84 k8s_client_request_latency_bucket{verb="GET",le="0.01"} 86 k8s_client_request_latency_bucket{verb="GET",le="0.025"} 89 k8s_client_request_latency_bucket{verb="GET",le="0.05"} 99 k8s_client_request_latency_bucket{verb="GET",le="0.1"} 112 k8s_client_request_latency_bucket{verb="GET",le="0.25"} 117 k8s_client_request_latency_bucket{verb="GET",le="0.5"} 117 k8s_client_request_latency_bucket{verb="GET",le="1"} 117 k8s_client_request_latency_bucket{verb="GET",le="2.5"} 117 k8s_client_request_latency_bucket{verb="GET",le="5"} 117 k8s_client_request_latency_bucket{verb="GET",le="10"} 117 k8s_client_request_latency_bucket{verb="GET",le="+Inf"} 117 k8s_client_request_latency_sum{verb="GET"} 2.1254330859999997 k8s_client_request_latency_count{verb="GET"} 117 k8s_client_request_latency_bucket{verb="POST",le="0.005"} 5 k8s_client_request_latency_bucket{verb="POST",le="0.01"} 5 k8s_client_request_latency_bucket{verb="POST",le="0.025"} 5 k8s_client_request_latency_bucket{verb="POST",le="0.05"} 6 k8s_client_request_latency_bucket{verb="POST",le="0.1"} 6 k8s_client_request_latency_bucket{verb="POST",le="0.25"} 6 k8s_client_request_latency_bucket{verb="POST",le="0.5"} 6 k8s_client_request_latency_bucket{verb="POST",le="1"} 6 k8s_client_request_latency_bucket{verb="POST",le="2.5"} 6 k8s_client_request_latency_bucket{verb="POST",le="5"} 6 k8s_client_request_latency_bucket{verb="POST",le="10"} 6 k8s_client_request_latency_bucket{verb="POST",le="+Inf"} 6 k8s_client_request_latency_sum{verb="POST"} 0.048558582 k8s_client_request_latency_count{verb="POST"} 6 k8s_client_request_latency_bucket{verb="PUT",le="0.005"} 1 k8s_client_request_latency_bucket{verb="PUT",le="0.01"} 1 k8s_client_request_latency_bucket{verb="PUT",le="0.025"} 1 k8s_client_request_latency_bucket{verb="PUT",le="0.05"} 1 k8s_client_request_latency_bucket{verb="PUT",le="0.1"} 1 k8s_client_request_latency_bucket{verb="PUT",le="0.25"} 1 k8s_client_request_latency_bucket{verb="PUT",le="0.5"} 1 k8s_client_request_latency_bucket{verb="PUT",le="1"} 1 k8s_client_request_latency_bucket{verb="PUT",le="2.5"} 1 k8s_client_request_latency_bucket{verb="PUT",le="5"} 1 k8s_client_request_latency_bucket{verb="PUT",le="10"} 1 k8s_client_request_latency_bucket{verb="PUT",le="+Inf"} 1 k8s_client_request_latency_sum{verb="PUT"} 0.002381375 k8s_client_request_latency_count{verb="PUT"} 1 k8s_client_request_total{code="200",method="GET"} 120 k8s_client_request_total{code="200",method="PUT"} 1 k8s_client_request_total{code="409",method="POST"} 6 ``` Signed-off-by: Andrew Dye <[email protected]> * Histogram Bucket Options Add abstraction to be able to pass buckets custom defined to histogram vectors. Signed-off-by: Andrew Dye <[email protected]> * Add org to CreateUploadLocation Signed-off-by: Andrew Dye <[email protected]> * Add config for grpc MaxMessageSizeBytes We need to make the grpc max recv message size in propeller's admin client configurable to match the server-side configuration we support in admin. Signed-off-by: Andrew Dye <[email protected]> * Move storage cache settings to correct location Signed-off-by: Andrew Dye <[email protected]> * added lock to memstore make threadsafe Signed-off-by: Andrew Dye <[email protected]> * Add read replica host config and connection - Add a new field to the postgres db config struct, `readReplicaHost`. - Add a new endpoint in the `database` package to enable establishing a connection with a db without creating it if it doesn't exist Signed-off-by: Andrew Dye <[email protected]> * Fix type assertion when an event is missed while connection to apiser… …ver was severed Signed-off-by: Andrew Dye <[email protected]> * Log and monitor failures to validate access tokens Signed-off-by: Andrew Dye <[email protected]> * Dask dashboard should have a separate log config Signed-off-by: Andrew Dye <[email protected]> * adjust Dask LogName to (Dask Runner Logs) Signed-off-by: Andrew Dye <[email protected]> * Fix k3d local setup prefix I was trying to use `setup_local_dev.sh`, and it wasn't working out of the box. Looks like it expects `k3d-` prefix for the kubecontext Ran `setup_local_dev.sh` Signed-off-by: Andrew Dye <[email protected]> * Override ArrayNode log links with map plugin This PR adds a configuration option to override ArrayNode log links with those defined in the map plugin. The map plugin contains it's own configuration for log links, which may differ from those defined on the PodPlugin. ArrayNode, executing subNodes as regular tasks (ie. using the PodPlugin) means that it uses the default PodPlugin log templates. Signed-off-by: Andrew Dye <[email protected]> * Add histogram stopwatch to stow storage This change * Adds a new `HistogramStopWatch` to promutils. This [allows for aggregating latencies](https://prometheus.io/docs/practices/histograms/#quantiles) across pods and computing quantiles at query time * Adds `HistogramStopWatch` latency metrics for stow so that we can reason about storage latencies in aggregate. Existing latency metrics remain. - [x] Added unittests Signed-off-by: Andrew Dye <[email protected]> * Fix metrics scale division in timer * Fix metrics scale division in timer Signed-off-by: Iaroslav Ciupin <[email protected]> Signed-off-by: Andrew Dye <[email protected]> * CreateDownloadLink: Head before signing Signed-off-by: Andrew Dye <[email protected]> * Unexpectedly deleted pod metrics * Count when we see unexpectedly terminated pods Signed-off-by: Andrew Dye <[email protected]> * Don't send inputURI for start-node * send empty `inputUri` for `start-node` in node execution event to flyteadmin and therefore, GetNodeExecutionData will not attempt to download non-existing inputUri as it was doing before this change. * add DB migration to clear `input_uri` in existing `node_executions` table for start nodes. Signed-off-by: Andrew Dye <[email protected]> * Fix cluster pool assignment validation Signed-off-by: Andrew Dye <[email protected]> --------- Signed-off-by: Andrew Dye <[email protected]> Signed-off-by: Eduardo Apolinario <[email protected]> Co-authored-by: Dan Rammer <[email protected]> Co-authored-by: Joe Eschen <[email protected]> Co-authored-by: Katrina Rogan <[email protected]> Co-authored-by: Michael Barrientos <[email protected]> Co-authored-by: Haytham Abuelfutuh <[email protected]> Co-authored-by: Jan Fiedler <[email protected]> Co-authored-by: Iaroslav Ciupin <[email protected]> Co-authored-by: Eduardo Apolinario <[email protected]>
- Loading branch information
9 people
authored
Oct 22, 2024
1 parent
71386f8
commit e4d29f8