[KubeCon EU 2020] containerd Deep Dive

containerd Deep Dive
Akihiro Suda (NTT) & Wei Fu (Alibaba Cloud)
KubeCon EU 2020 Virtual (Aug 19, 2020)

containerd overview
Akihiro Suda (NTT)

What is ?
● “Mid-level Container runtime”
○ Below platforms (Docker, Kubernetes)
○ Above lower level runtimes (runc)
● Resource Manager
○ Container processes
○ Image artifacts
○ Filesystem snapshots
○ Metadata and dependencies
● CNCF graduated project since February 2019
○ Following Kubernetes, Prometheus, Envoy,
and CoreDNS

Highly customizable
● Runtime plugins
○ Runc, gVisor, Kata, Firecracker...
● Snapshotter plugins
○ OverlayFS, BtrFS, ZFS, …
● Content store plugins
○ Local, IPFS...
● Stream processor plugins
○ ImgCrypt, zstd...

Adoption of containerd
● Container engines
● Kubernetes distributions
● Managed Kubernetes Services
Docker & Moby k3c PouchContainer
k3s kubespray microk8s
Alibaba ACK
Amazon EKS
(Fargate nodes)
Azure AKS
Google GKE IBM IKS
kind minikube
Charmed
Kubernetes
And more...

Adoption of containerd
● BuildKit
○ The modern implementation of `docker build`
● LinuxKit
○ Small Linux distro with containerd as the init
● Faasd
○ OpenFaaS for containerd
● VMware Fusion Nautilus
○ containerd on macOS, using VMware as the runtime plugin

Upcoming features in v1.4
Akihiro Suda (NTT)

Lazy pulling of images
● Run containers before completion of downloading the images
● Use cases:
○ Python/Ruby/Java/dotNET images
○ FaaS
○ Web apps with huge amount of HTML templates and media ﬁles
○ Jupyter Notebooks with big data samples included
○ Full GNOME/KDE desktop

Lazy pulling of images: Stargz & eStargz
● The containerd snapshotter plugin for Stargz & eStargz
https://github.com/containerd/stargz-snapshotter
● Stargz: seekable tar.gz for lazy-pullable container images
● eStargz: extended Stargz for batching frequently used ﬁles
● Both are fully compatible with legacy OCI tar.gz

Metadata 0
File 0
Metadata 1
File 1
Metadata {n-1}
File {n-1}
Footer
...
gzip
legacy tar.gz Stargz
Metadata 0
File 0
gzip
Metadata 1
File 1
gzip
...
Metadata {n-1}
File {n-1}
gzip
gzip
Footer
Metadata
stargz.index.json
Can’t inspect file offsets without
reading the whole archive
Can inspect the file offsets
immediately

● eStargz profiles the actual file access pattern and reorders the file entries,
so that relevant files can be prefetched in a single HTTP request
/usr/bin/apt-get
/bin/ls
/bin/vi
/lib/libc.so
/lib/libjpeg.so
/usr/bin/python3
.../usr/lib/python3/.../foo
/usr/lib/python3/.../bar
/app.py
/bin/ls
/app.py
/usr/bin/python3
/lib/libc.so
/usr/lib/python3/.../foo
/usr/lib/python3/.../bar
.../bin/vi
/lib/libjpeg.so
/usr/bin/apt-get
Stargz eStargz

Yesterday’s talk
https://sched.co/ZepQ

Support for SELinux MCS on CRI mode
● MCS: multi-category security
Containers
Volumes
UID=0
C42
UID=0
C42
UID=0
C43
UID=0
C43

Support for cgroup v2
● The new cgroup hierarchy, adopted by Fedora (since 31)
● Simpler layout
○ V1: /sys/fs/cgroup/{memory,cpu,devices,pids….}/foo
○ V2: /sys/fs/cgroup/foo
● Supports eBPF integration, pressure metrics, improved OOM control...
● Friendly to non-root users

Improved support for rootless mode
● Run containerd (and relevant components) as a non-root user
● Protect the host from potential vulnerabilities
● Adoption in containerd-related projects
○ Docker
○ BuildKit
○ k3s
○ k3c (on plan)
○ Kubernetes (on proposal, KEP 1371)

● [v1.3] No support for resource limitation (docker run --cpus … --memory ...)
○ Because unprivileged users cannot control cgroups
● [v1.3] No support for overlayfs snapshotter
○ Because unprivileged users cannot mount overlayfs
(except on Ubuntu/Debian kernels)
○ “Native” snapshotter can be used, but slow and wastes the disk

● [v1.3] No support for resource limitation (docker run --cpus … --memory ...)
○ Because unprivileged users cannot control cgroups
● [v1.3] No support for overlayfs snapshotter
○ Because unprivileged users cannot mount overlayfs
(except on Ubuntu/Debian kernels)
○ “Native” snapshotter can be used, but slow and wastes the disk
→ v1.4 supports resource limitation
(requires cgroup v2 and systemd)
→ v1.4 supports FUSE-OverlayFS snapshotter
(requires kernel >= 4.18)

Demo: Rootless Kubernetes with Cgroup v2
“Usernetes” https://github.com/rootless-containers/usernetes
https://asciinema.org/a/349859

Other changes in v1.4
● Windows CRI
● systemd NOTIFY_SOCKET
● Support reloading CNI conﬁg without restarting the daemon
● Socat binary is no longer needed
Release note: https://github.com/containerd/containerd/releases

v1.5 planning
● NRI: Node Resource Interface (#4411)
○ The new common interface for node resources such as cgroup
○ The plugin spec is very similar to CNI
● Sandbox API (#4131)
○ Pod sandbox as a ﬁrst-class object
○ No “/pause” process
● Filesystem quota (#759)

containerd: external plugins
Wei Fu (Alibaba Cloud)

Backend as external plugins
● Big goal - no re-compilation required!!!
● Stream processors
● gRPC proxy plugin for image storage
● RuntimeV2 proto for OCI Runtime

Stream processor
● OCI Image layer data packaged in tar archive
● OCI image spec only supports few compression algorithms
○ +gzip/+zstd, but +gzip is more common
● How to handle experimental media-type stream?
○ Or encryption purpose?
Image
Layer
Snapshot
Tar Stream
Processor
Diff Service
+gzip
Custom?

Stream processor
● Stream processor(SP) is binary plugin handling media-type stream
○ Accepts customize media-types, returns other one
○ Call binary for media-type converter
● Example
○ containerd/imgcrypt
Image
Layer
Snapshot
Tar
SP
Diff Service
Tar+Gzip
SP
Tar(+Gzip)+encrypted
SP
Other Customize SP

Stream processor - Demo
● Integrate with +zstd media-type
● asciinema link
[stream_processors]
[stream_processors."zstd"]
accepts = ["application/vnd.oci.image.layer.v1.tar+zstd"]
returns = "application/vnd.oci.image.layer.v1.tar"
path = "zstd"
args = ["-dcf"]

Snapshot proxy plugin
// Snapshot service manages snapshots
service Snapshots {
rpc Prepare(PrepareSnapshotRequest) returns (PrepareSnapshotResponse);
rpc View(ViewSnapshotRequest) returns (ViewSnapshotResponse);
rpc Mounts(MountsRequest) returns (MountsResponse);
rpc Commit(CommitSnapshotRequest) returns (google.protobuf.Empty);
rpc Remove(RemoveSnapshotRequest) returns (google.protobuf.Empty);
rpc Stat(StatSnapshotRequest) returns (StatSnapshotResponse);
rpc Update(UpdateSnapshotRequest) returns (UpdateSnapshotResponse);
rpc List(ListSnapshotsRequest) returns (stream ListSnapshotsResponse);
rpc Usage(UsageRequest) returns (UsageResponse);
}

Snapshot proxy plugin
package main
import(
"net"
"log"
"github.com/containerd/containerd/api/services/snapshots/v1"
"github.com/containerd/containerd/contrib/snapshotservice"
)
func main() {
rpc := grpc.NewServer()
sn := CustomSnapshotter()
service := snapshotservice.FromSnapshotter(sn)
snapshots.RegisterSnapshotsServer(rpc, service)
// Listen and serve
l, err := net.Listen("unix", "/var/run/mysnapshotter.sock")
if err != nil {
log.Fatalf("error: %vn", err)
}
if err := rpc.Serve(l); err != nil {
log.Fatalf("error: %vn", err)
}
}
● Conﬁgure with proxy_plugins
● Example
○ stargz-snapshotter
○ CVMFS Containerd Snapshotter
[proxy_plugins]
[proxy_plugins.customsnapshot]
type = "snapshot"
address = "/var/run/mysnapshotter.sock"

Runtime V2
● A ﬁrst class shim API for runtime authors to integrate with containerd
○ More VM like runtimes have internal state and more abstract actions
○ A CLI approach introduces issues with state management
○ Each runtimes has its own values, but keep containerd in solid core scope
● Example
○ gVisor
○ KataContainer
○ Firecracker

Runtime V2
service Task {
rpc State(StateRequest) returns (StateResponse);
rpc Create(CreateTaskRequest) returns (CreateTaskResponse);
rpc Start(StartRequest) returns (StartResponse);
rpc Delete(DeleteRequest) returns (DeleteResponse);
rpc Pids(PidsRequest) returns (PidsResponse);
rpc Pause(PauseRequest) returns (google.protobuf.Empty);
rpc Resume(ResumeRequest) returns (google.protobuf.Empty);
rpc Checkpoint(CheckpointTaskRequest) returns (google.protobuf.Empty);
rpc Kill(KillRequest) returns (google.protobuf.Empty);
rpc Exec(ExecProcessRequest) returns (google.protobuf.Empty);
rpc ResizePty(ResizePtyRequest) returns (google.protobuf.Empty);
rpc CloseIO(CloseIORequest) returns (google.protobuf.Empty);
rpc Update(UpdateTaskRequest) returns (google.protobuf.Empty);
rpc Wait(WaitRequest) returns (WaitResponse);
rpc Stats(StatsRequest) returns (StatsResponse);
rpc Connect(ConnectRequest) returns (ConnectResponse);
rpc Shutdown(ShutdownRequest) returns (google.protobuf.Empty);
}

Runtime V2 - Binary
● Binary naming convention
○ Name io.containerd.runc.v2 --> Binary containerd-shim-runc-v2
■ So both io.containerd.runc.v1 and io.containerd.runc.v2 are runtime V2
■ runc.v2 supports grouping several containers with less resource
■ runc.v2 as CRI plugin’s default runtime
○ Via a runtime binary available in containerd’s PATH
● Required start/delete sub-commands
○ Resources created by container will be cleanup by delete sub-command

Runtime V2 - Logging
● ﬁfo/npipe as default channel
○ Receiver consumes more resources to handle log output.
dockerd
CRI-plugin
containerd shim
kernel
Named Pipe

● ﬁfo/npipe as default channel
○ Receiver consumes more resources to handle log output.
○ And it requires that receiver must be alive!!!
○ Impact running containers if receiver is down too long.
containerd shim
Named Pipe
kernel

● Support pluggable logging via STDIO URIs
○ fifo - Linux (default)
○ npipe - Windows (default)
○ binary - Linux & Windows
○ file - Linux & Windows
schema path:// ?key=valueSTDIO URI
file file :// /var/log/cntr/hi ?maxSize=100MB
binary binary :// /usr/bin/syslog ?addr=192.168.0.3

[KubeCon EU 2020] containerd Deep Dive

More Related Content

What's hot

What's hot (20)

Similar to [KubeCon EU 2020] containerd Deep Dive

Similar to [KubeCon EU 2020] containerd Deep Dive (20)

More from Akihiro Suda

More from Akihiro Suda (20)

Recently uploaded

Recently uploaded (20)

[KubeCon EU 2020] containerd Deep Dive