Status

Current state: [ UNDER DISCUSSION ]

Discussion thread: <link to mailing list DISCUSS thread>

JIRA: SAMZA-TBD

Released: 

Problem

Samza operates in a multi-tenant environment with cluster managers like Yarn and Mesos where a single host can run multiple Samza containers. Often due to soft limits configured for cluster managers like Yarn and no notion of dynamic workload balancing in Samza a host lands in a situation where it is underperforming and it is desired to move one or more containers from that host to other hosts. Today this is not possible without affecting other jobs on the hot host or restarting the affected job manually. In other use cases like resetting checkpoints of a single container or supporting canary or rolling bounces the ability to restart a single or a subset of containers without restarting the whole job is highly desirable. 

This doc addresses the problem of restarting/moving a single container of a job without affecting other containers of the same job or other job running on the same host. X 

Motivation

Alleviating Hot Host Problems: Yarn as a resource manager in Samza is configured to operate on soft limits (a job sets vcore for a container or has defaulted, this is the minimum guarantee of resources to be provided by Yarn ) & most of the customers go by default and fail to right-size their Job. In addition, there is no notion of a dynamic cluster balancer in Samza at LinkedIn today. Although Yarn to some capacity acts as a Cluster Balancer if configured with hard limits. Due to this often a host lands in a situation when containers on it are underperforming (also referred to as a hot host) because it has CPU heavy containers running on it while some other hosts are underutilized. Now it is desirable to move this container to a different host. To achieve the same following solutions exists:

1. Rewrite the locality mapping in coordinator stream and restart the job 
2. Take the hot host out of rotation which kills containers from all the jobs running on that host. Then trigger a restart on a job whose container is supposed to be moved so Yarn would try to allocate some other host for it. Once the container starts on other hosts, then put the hot host in rotation again so that other container who were killed as a result of taking host out can be attempted to restart on the hot host again. It ain't easy!

If the ability to move a container exists, someone at the simplest can manually move containers to different hosts or write some simple scripts to automate that.

Canary / Rolling Bounces: When there is a bug in the Samza framework code that affects the Samza container deployment, Samza engineer needs to manually restart the container process on the given machine with the given binary version. This involves multiple steps (e.g. manually identify and log in the container host before using kill -9 to stop the process) which is inconvenient. With the restart ability, the same system can be used for building support for Canary or Rolling Bounces for YARN based Samza deployments, restart ability can be easily extended to deploy a single or subset of containers using a different version of application code. 

Resetting Checkpoints: Startpoint API has made resetting checkpoints easy but it still needs a dev to restart his job once he has set his start points. The restart can be potentially be prevented with restart ability of a single container or a few containers

Draining a host: Moving all running containers from a host sequentially to other hosts or in parallel.

Fix a Job in Degraded State: Often users find it desirable to just restart a single container for various reasons like underperforming containers when only a few containers running into exceptions because some partitions have corrupt messages. Today Samza kills the job if a container has run into exceptions more than a fixed configured number of times. In these scenarios, it's desirable for Users to keep the job running in a degraded state and only fix one or few containers of the job and issue a restart for them. 

Dynamic Workload Balancer: The ability to move and restart containers is the fundamental building block to developing a load balancer (like Cruise Control) for Samza. At the very simple this load balancer can be a simple script trying to balance cluster, later it can be built into a more sophisticated system. 

Heterogeneous container: Each Samza job today has homogenous containers (in regards to memory & vcore configurations). One of the desired use cases for Samza in the future is the ability to restart a container with different sizes (memory & cpu) which can be used by ay auto-sizing engine designed for samza.

Selectively Spin StandBy Containers: Samza has a feature of Hot StandBy Containers for reducing stateful restoration times. Enabling this feature for a job involves doubling the containers at the least (simplest case where every container has 1 standby replica enabled). Customers are reluctant to enable this since doubling the containers increases the cost to serve. To improve the adoption for this feature we can build the ability to spin up StandBy Containers for a single or a subset of containers while the job is running, these StandBy Containers then can be used for failover to reduce downtime.  

Requirements

Goals

• Build the ability to move container (non-am) of a running job from its existing host to a preferred host (if specified) or any host
• If an existing stateful container has a standby container, move to the active container to stand by container 
• If an existing stateful container is without a standby container, spin up a standby container and then move the active to standBy once stand by has bootstrapped to reduce downtime [Stretch]
• Build the ability to restart containers with the same host preferences
• Build the ability to selectively spin up standby for one or a subset of containers [Stretch]
• Expose these APIs via a tool or dashboard for Samza team to easily use during oncall 

Non-Goals

• This system does not intend to solve the problem of Dynamic Workload Balance i.e  Cruise control. It may act as a building block for one later.
• Solving canary problem for YARN based deployment mode is out of scope of this solution however system built should be easily extensible to support canary 
• This system will not have a built in intelligence to find a better match for host for a container it will take simplistic decisions per params passed by user.

SLA / SCALE / LIMITS

• At a time AM for a single job will only serve one request per container, parallel requests across containers are still supported. If a control request is underway any other requests issued on the same container will be discarded
• System should be capable of scaling to be used across different jobs at the same time 

General Requirements

At a high level the system can be divided into two logical components:

1. Container Placement Handler (CPH): to serve as a dispatcher of specific control requests from a control plane. Fetches the container placement control action from control-plane, queues it up and dispatches requests on containers of a job (non-am) with policy (details below) 
2. Container Placement Service: API layer built around Cluster Manager to move containers. Reacts to control actions from CPH and interacts with underlying Cluster manager to execute them 

• Requirements for Container Placement Handler

1. Should be able to handle multiple move request for different containers at the time for a single job
2. Should provide the good monitoring mechanism 
3. Should be general enough to be coupled with any system for example Control-plane
4. Should allow queuing requests for multiple containers across multiple jobs at the same time 
5. Should be easily extensible to build canary support for yarn deployments 

• Requirements for Container Placement Service

1. Under no circumstances should Control Actions result in an inconsistent / non-working state of the job (i.e., #running-containers should equal number of configured containers).
2. Should not be tightly coupled with YARN, should be extensible to be used by any cluster manager implementation

Failure Scenarios & Assumptions

1. Each control action will be associated with a deployment id of an app it is intended to be issued upon
2. Across Job restarts, any requests issued for the previous deployment incarnation will not be respected. This includes AM restarts (today if the AM restarts that implies a job restart)

Glossary

Proposed Changes

API design

On the basis of types of Control actions, the commands are the following:       

1. Restart Container

Note: For supporting canary above parameter list can be easily extended to support following parameters

3. Container Status

2. Enable & Disable StandBy [Stretch]

Architecture

For implementing a scalable container placement control system, the proposed solution is divided into two parts:

Part 1. Container Placement Handler

1. Control Plane as described plane above the job that allows taking control actions by multiple controllers like Samza Dashboard, Start points controller. 
2. ContainerPlacementHandler is handler registered to control plane that translates control actions to invoke Container Placement Service APIs

This control plane can be implemented in the following ways 

Option 1: Samza metastore service [Preferred]

Samza Metastore will provide an API to write to the coordinator stream. One simple way to expose Container Placement API is, Container Placement handler can have a coordinator stream consumer polling control messages from coordinator stream & acting on them. CPH will take actions maintaining some in-memory state with Container Placement Service to not take an action twice. In addition since control actions are associated with a deployment id, they are automatically invalidated across restarts 

Option 2: Stub a Rest Endpoint in JC [Rejected Alternative]

It's fairly simple to embed a REST endpoint in JC or extend the existing Rest endpoint in JC that exposes configs in JC to support apis listed above 

Open Sources Access: 

To expose this API to open source users, a metadata store writer tool can be used. The same tool is going to give Open source users access to Start points APIs.

Part 2. Container Placement Service

Container Placement service is a set of API Built around AM to move/restart containers. The solution proposes to introduce a ContainerPlacementManager which is a single entity managing the container placements for stateful and stateless jobs & a metadata holder for container actions (ControlActionMetadata) forex request_id, current status, requested resources, etc. 

2.1 Container Move

2.1.1 Stateless Container Move (Host Affinity Off)

Option 1: Reserve-Stop-Move: Request Resource first and Only make a move once resources are allocated by ClusterResource Manager (Preferred)

Option 2: Stop-Reserve-Move: Stop the container and then make resource request for preferred resources and attempt a start (Rejected Alternative)

Orchestration of Stateless Container Move using Option 1: The image shows a sequence of steps involved in moving a container. Note that there is a specific number of retries configured for requesting preferred hosts to ClusterResourceManager.

1. ContainerPlacementHandler dispatches the move request to ContainerProcessManager 
2. ContainerProcessManager registers move with ContainerPlacementManager and issues a resource request with ContainerAllocator (Async)
3. Now there are two possible scenarios

3.1 Requested Resources are allocated by ClusterResourceManager 

1. In this case ContainerPlacementManager initiates a container shutdown and waits for the ContainerPlacementManager callback on successfully stopping the container
2. ContainerProcessManager on receiving the callback checks if this was due to a move request & propagates it to ContainerPlacementManager which issues a start container request ClusterResourceManager to start container on allocated resources 
3. If the container start request succeeds then a success notification is sent (updating the Control Action Metadata) otherwise a new request to start this container on ANY_HOST is issued

3.2 Resource Request expires

1. In cases where Resource Request expires and ClusterResourceManager is not able to allocate requested resources there a fixed number times a request is retried. If request does not get allocated even after retires, move is marked as failed and container stays on the host it is running on. 

Failure Scenarios:

• If the ContainerPlacementManager is not able to stop the active container (3.1 #1 above fails) in that case we mark the control action request failed
• If ClusterResourceManager fails to start the stopped active container on accrued destination host, then we attempt to start the container back on source host, failing to do which will attempt to start the container on ANY_HOST   

2.1.2 Stateful Container Move (Host Affinity On)

The stateful container is divided into two cases on the basis of whether the job has StandBy Container for it or not

Case 1. Stateful Job has Standby Container Enabled

Let's take a case when the job has a stateful container C on H1 and has a StandBy container enabled C’ on H2

Option 1: Stop-Reserve-Move: Initiate a Stop on C & issue a Standby failover (Preferred - Phase 1*)

In this option, C is stopped on H1 and failover is initiated from H1 to H2 with current semantics of Failover of an active Container in the Hot standby feature . This ensures C moves either to H2 (in the best case) or to some other hosts (when Cluster Manager cannot allocate capacity on H2, i.e H2 is under capacity)

Option 2: Reserve-Stop-Move: Initiate a Standby move first then issue a move for Active Container (Preferred - Phase 2*)

1. In this option request is first issued to stop C’ 
2. Then request H2 from cluster manager to start C 
3. Issue a stop on C only H2 can be allocated (#2 succeeds). 
4. Then request C’ to start ANY_HOST except for H1, H2 

Two ways to achieve this: 

• A client can make two calls for this first move C’ to ANY_HOST apart from H2, H1 (similar to a stateless move), then move the request of C to C’ (similar to a stateless move) so the state maintenance of this lives on the client
• Maintain state in JC to accomplish this 

* Phase 1 & Phase 2 refer to the implementation phases, please see the section: Implementation Plan 

Option 3: Reserve-Stop-Move: First Request H2 from Cluster Manager and initiate the failover once you get H2 as a preferred resource (Rejected Alternative)

• In this option C’ is stopped on H2 and failover is initiated from H1 to H2 with current semantics as developed by the feature Hot standby
• In case of the move failed only affects standby

Orchestration of StandBy Enabled Container Move using Option 1 (Phase 1)

1. ContainerPlacementHandler dispatches the move request to ContainerProcessManager
2. ContainerProcessManager registers a move & initiates a failover with StandByContainerManager then the following failover sequence is followed

Case 2. Stateful Job has Standby Container Disabled [Stretch]

Option 1: Stateful containers move is equivalent to Stateless container Move (Phase 1)

In this option, Stateful container move like stateless container & does a bootstrap

Option 2: Optionally Spin Up StandBy container & then move (Phase 2)

If we expose an API to spin a StandBy Container for an active container, a client can then make two requests, one to spin up a StandBy then periodically monitor the lag for StandByCotntainer (Using Diagnostics). Once the StandBy Container has a steady lag is ready to move, this case becomes Case 1 mentioned above. (More details TBD later)

2.2 Stateful & Stateless Container Restart

This is much simpler as compared to Move. We can either make restart equivalent to a move or issue a stop a container and ContainerProcessManager will try to start that container again. Let's discuss these options in detail:

Option 1: Reserve-Stop-Start: Restart is equivalent to move on the same host [Preferred]

• In this option, resources are requested on the same host first before issuing a stop on Container
• Once the resources are accrued a container is stopped & requested to start 
• So this is equivalent to move semantics above 

Both stateless & stateful container restarts will be equivalent to Stateless Container Move on the same host. Since the container is restarted on the same host then

Option 2: Stop-Reserve-Start: Restart is equivalent to stopping container first then attempting to start it [Rejected Alternative]

• In this option, a container is issued a stop first
• Once the container stops, resources are requested for starting on the same host (last seen host)
• Once the resources are accrued, the container is issued a start on the same host

Case 2.1: Host Affinity is OFF 

In this case, when User registers a restart request, ContainerPlacementManager issues a stop on the current container. On receiving a callback from ClusterResourceManager for container stop, ContainerProcessManager requests resources on last seen host. But Allocator thread assigns this request on ANY_HOST available since we have homogenous containers in Yarn & host affinity is off, the container can be started from any available resource.

Case 2.2: Host Affinity is ON 

In this case, a Stop is issued on Container & the ContainerProcessManager makes a request for resources on last seen host to ClusterResourcesManager. Allocator Thread on having resources on the preferred host (here last seen host) issues a request to run a container on it. If this request expires then there are multiple possibilities: 

• If StandBy Containers are enabled directly initiate a failover 
• Else if there are resources on ANY_HOST then issues a container start request on that resource
• Else issue an ANY_HOST request to ClusterResourceManager

Public Interfaces

Implementation and Test Plan

Implementation Plan

Test Specifications

• Complete testing plan: https://docs.google.com/spreadsheets/d/1v-fw0pHxKRobGkALDCno4FuPCsBhdepQ86vIGLHWu54/edit?usp=drive_web&ouid=115242259601904072922
• Local deployment testing with Virtual Private Cluster (LXC with Samza)
• Create a test job using VPC or in the real cluster that automates testing for taking control actions on containers

Compatibility, Deprecation, and Migration Plan

• The new interfaces & API introduced should not affect any part of the User code & should be backward compatible. No migration needed for this new feature.
• Jobs using older samza version shall be able to discard the Control messages written to the metastore

Rejected Alternatives

Note: These are described in the implementation along with preferred options above for the sake of consistency in reading

References

1. https://docs.google.com/document/d/18EdLZgrW-ZMIypqI2YX0YsAnTa0Zf_0Rt4tQxkLvZ10/edit#
2. https://www.cloudera.com/documentation/enterprise/5-10-x/topics/cm_mc_rolling_restart.html
3. https://docs.google.com/document/d/16cQHipnAZHCerEgRW6icw9YbWA3TUcZ2bm-g9D79y6c/edit#