1.
ScyllaDB Cloud Goes
Serverless
Yaniv Kaul, VP R&D, ScyllaDB

2.
Yaniv Kaul
■ Joined ScyllaDB in October 2022
■ Previously, at Red Hat, managed
■ Storage (Ceph, cloud storage),
■ Virtualization (OpenStack)
■ Performance and Scale.
■ Years ago, worked with ScyllaDB founders on the KVM
hypervisor

3.
■ Introduction
■ Tour
■ Architecture
Presentation Agenda

4.
Introduction

5.
ScyllaDB Cloud
■ Our managed ScyllaDB clusters offering
■ Provides clusters running on AWS or GCP clouds
■ Uses whole virtual machine(s) per tenant
■ Use our infrastructure or ‘bring-your-own-account’ nodes
■ Lifecycle, monitoring, logging, alerting - overall management by our DevOps and
Support teams
■ Free trial also available!
cloud.scylladb.com

6.
ScyllaDB Cloud - K8S based
The next generation ScyllaDB Cloud is moving to run ScyllaDB workloads on Kubernetes (K8S).
■ Infrastructure components (storage, network, security, monitoring, logging and more)
deployment, management are handled by Kubernetes
■ Provides seamless integration with cloud services
■ Based on the open source ScyllaDB K8S operator

7.
ScyllaDB Cloud - K8S based
Cluster Creation Phase

8.
ScyllaDB Cloud - K8S based
Cluster Is Up

9.
ScyllaDB Cloud - K8S based
General Tab

10.
ScyllaDB Cloud - K8S based
Connection Bundle

11.
ScyllaDB Cloud - K8S based
cqlsh Connection

12.
docker run -it --entrypoint cqlsh
-v
/file/downloaded/from/cloud/connect-bundle-ScyllaSummit2023.yaml
:/connect-bundle-ScyllaSummit2023.yaml:Z
scylladb/scylla-nightly:5.2.0-dev-0.20221215.16c50bed5eb8
--cloudconf connect-bundle-ScyllaSummit2023.yaml

13.
Python Example

14.
ScyllaDB Cloud - K8S based
Monitoring

15.
Architecture

16.
Challenges With K8S
ScyllaDB on K8S has some of the ‘usual’ challenges and some ‘unusual’ ones:
■ Stateful application (a database)
■ Strict requirements from the operating system
■ Deployment requirements

17.
The control plane is mostly running on a separate system
■ ScyllaDB Cloud management system was extended to treat serverless as
yet another cloud provider
■ Monitoring and logging are kept outside the operational K8S cluster running
the workloads
Management

18.
First, need to provision a K8S node (worker), that will be used to run tenants’
ScyllaDB nodes:
■ RAID(0) configuration, OS tuning, etc.
■ Disk IO benchmarking, to get the baseline performance numbers.
■ Avoid ‘lemons’
■ Distribute baseline IO performance between tenants (per tenant CPU
shares)
Data Plane / Workers

19.
The ScyllaDB K8S operator deploys a complete ScyllaDB cluster:
■ No. of nodes, size (CPU, memory, disk) , distribution, replication factor
(RF=3 for the time being), and other configuration items.
■ Configure / initialize node, wait for cluster formation
■ Manager (agent) & configuration
■ Monitoring
■ Drivers proxy (‘SNI proxy’)
Tenant Deployment

20.
■ Virtual machines forces a fixed cores/memory/disk relationship
■ Scaling can only be performed by adding or removing nodes
■ With K8S, we abstract underlying hardware and provide much more
flexibility.
Breaking the Fixed Compute Storage
Relationship

21.
■ Memory - each tenant (Pod) get their own (request/limit)
■ CPU - each tenant (Pod) get their own (request/limit)
■ Disk - each tenant get their own XFS mount point, with quota assigned
■ Scylla’s IO scheduler knows how much IOPS to give to each tenant
■ KMS is used to encrypt each tenant’s data
Tenant Isolation

22.
/mnt/md0
Tenant Isolation - Storage (quota)
Instance
disks
Host path XFS Quota
/mnt/md0/local-pv-1 50G
/mnt/md0/local-pv-2 100G
kind: ScyllaCluster
spec:
datacenter:
racks:
- storage:
capacity: 50G
kind: ScyllaCluster
spec:
datacenter:
racks:
- storage:
capacity: 100G
Dynamic
storage
provisioner
RAID 0
XFS

23.
Tenant Isolation - Storage (performance)
i3en.*xlarge
disks:
- mountpoint: /var/lib/scylla
read_iops: 125000
read_bandwidth: 1000000000
write_iops: 80000
write_bandwidth: 500000000
disks:
- mountpoint: /var/lib/scylla
read_iops: 125000
read_bandwidth: 1000000000
write_iops: 80000
write_bandwidth: 500000000
disks:
- mountpoint: /var/lib/scylla
read_iops: 250000
read_bandwidth: 2000000000
write_iops: 160000
write_bandwidth: 1000000000

24.
Performance
Example of workload testing:
■ 4 loaders
■ 7 tenants
■ 4 OLTP, 3 OLAP
workloads
■ Mixed stress workload
Workload OLTP OLTP OLTP OLTP OLAP OLAP OLAP
CPU cores per scylla node 2 / 16 2 / 16 2 / 16 2 / 16 2 / 16 2 / 16 2 / 16
Cassandra-stress stats Latency [ms] THROUGHPUT - OP/S
loader 1 - latency p99 read 2.2 2.2 2.4 2.2 5,729 6,263 6,033
loader 2 - latency p99 read 2.2 2.2 2.2 2.1 5,629 6,294 5,979
loader 3 - latency p99 read 2.2 2.2 2.3 2.2 5,690 6,291 5,986
loader 4 - latency p99 read 2.2 2.5 2.2 2.2 5,685 6,283 5,986
loader 1 - latency p99 write 3.8 4 4.1 4.1 8,735 9,076 9,149
loader 2 - latency p99 write 3.9 4 4 3.9 8,325 8,898 8,809
loader 3 - latency p99 write 3.9 4 4 4.2 8,455 8,905 8,819
loader 4 - latency p99 write 3.8 4.2 4 4.2 8,450 8,893 8,836
loader 1 - latency p99 mixed 4.1 4 4.3 4.1 4,815 5,005 5,127
loader 2 - latency p99 mixed 4.1 4 4.2 4.1 4,552 4,835 4,899
loader 3 - latency p99 mixed 4.1 4 4.2 4.1 4,589 4,835 4,906
loader 4 - latency p99 mixed 4.1 4.2 4.2 4.1 4,587 4,819 4,903

25.
The different CQL drivers connect to the cluster via a proxy, known as the SNI
(Server Name Indication) proxy:
■ Cluster data is stored in a YAML file, that clients use to connect to the
cluster.
■ Using TLS, the clients use the SNI field to indicate to which ScyllaDB node
they will connect to.
■ They still connect to all shards (shard-aware)
Connection to the Cluster

26.
TLS protocol extension, enables connecting to a specific hostname
What is SNI?

27.
As part of the ‘S301: ScyllaDB Operations’ course you can learn more about our K8S operator
and get some hands-on experience working with it.
ScyllaDB University Content

28.
Thank You
Stay in Touch
Yaniv Kaul
yaniv.kaul@scylladb.com
@YanivKaul
github.com/mykaul
www.linkedin.com/in/ykaul/