BigQuery の内部実装の変遷について
BigQuery(正確にはそのクエリエンジンである Dremel)の内部実装の変遷をまとめた以下のブログポストおよび論文を読みました。
とても面白い内容で、Twitter にメモをポストしたのですが、後で参照しやすいようにブログにも同じ内容を載せておきます。
BigQuery(正確にはそのクエリエンジンである Dremel)の内部実装の変遷をまとめた論文の解説ブログ。面白かった。https://t.co/s9UaH4MjBS
— south37/Nao Minami (@south37777) December 28, 2020
解説対象の論文は "Dremel: A Decade of Interactive SQL Analysis at Web Scale"。上記ブログポストでは 3 章の Disaggregation と 5 章の Serverless Computing をピックアップして解説していたが、より網羅的に知りたい場合 & 原文を読みたい場合はこちらを読むと良い。https://t.co/XolZ7fMXFY
— south37/Nao Minami (@south37777) December 28, 2020
以下、雑なメモ。まず前提として 2010 年の Dremel についての最初の論文である "Dremel: Interactive Analysis of Web-Scale Datasets" の 6 章 Query Execution は目を通しておくと良い。Dremel が Query をどう実行するかが説明されている。https://t.co/A98EC7oBqE
— south37/Nao Minami (@south37777) 2020年12月28日
Dremel は「巨大なデータに対しての解析 query」である入力 query を「小さなデータに対しての解析 query」の集合に分解する。分解は 2010 年時点では tree の形で行われて、leaf query は storage からデータを読み取る形で実行される。中間 node は「child から集めたデータに対する query」を実行。
— south37/Nao Minami (@south37777) 2020年12月28日
中間 node の挙動は「query の書き換え」として実現される。具体的には、"SELECT A, COUNT(B) FROM T GROUP BY A" は "SELECT A, SUM(c) FROM (R1 UNION ALL ... Rn) GROUP BY A" へと書き換えられる。Ri は child node で query を実行した結果となっている。
— south37/Nao Minami (@south37777) 2020年12月28日
上記の Query Execution の前提がある中で、Dremel が 2010 年時点から進化したポイントがいくつかある。1つは、storage として local storage ではなく GFS (やその後継の Colossus)を利用するようになったこと(= Disaggregated storage)。
— south37/Nao Minami (@south37777) 2020年12月28日
元々 Dremel の server(特に leaf server)は local storage を利用していた。それが GFS を利用するようになり、Dremel の信頼性向上やメンテナンス性向上につながった(fully-managed な GFS を利用することで robustness が向上したり、local disk へのデータロードが不要になったりしたらしい)
— south37/Nao Minami (@south37777) 2020年12月28日
Dremel は join, aggregation, analytic operation のために shuffle という機構を持つが、それも進化したらしい。これは以下の blog post でも説明されてる。shuffle は元々 Map Reduce で導入され、Hadoop, Spark など主要な分散データ処理システムの動作を支える重要な機構https://t.co/ycJwG0aTo4
— south37/Nao Minami (@south37777) 2020年12月28日
2014 年に Dremel の shuffle は進化して、中間データを " distributed transient storage system" に持つようになった(= Disaggregated memory)。これにより、完全な in-memory での query 実行が実現されるようになったらしい。
— south37/Nao Minami (@south37777) 2020年12月28日
新しい shuffle infrastructure は latency 削減や shuffle 対象のデータ量の向上、20%以上の resource cost 削減に繋がった(おそらく、resource utilization が高まったのだと思われる)とのこと。ちなみに、同じ shuffle infrastructure が Google Cloud Dataflow でも使われているらしい。
— south37/Nao Minami (@south37777) 2020年12月28日
また、disaggregated memory shuffle system が利用されるようになったことで、query 実行の compute resource の割り当て(= scheduling)に柔軟性が生まれた。具体的には、shuffle 結果を checkpoint として、動的に worker を preempt 出来るようになったらしい(= Shuffle Persistence Layer)
— south37/Nao Minami (@south37777) 2020年12月28日
また、2010年時点では "fixed execution tree" として実行されていた query が、スケジューラの進化(= "centralized scheduling")と "shuffle persistence layer" のおかげで「DAG として表現された query plan に対して、柔軟に worker を割り当てる」という形で実行されるようになった。
— south37/Nao Minami (@south37777) 2020年12月28日
まとめると、storage, memory など様々な resource が分離され、それによって compute resource の柔軟な割り当てが可能になった。performance, resource utilization の両面で劇的に進化しており、さらに改善が積み重なっている様子が感じ取られて面白い内容だった。
— south37/Nao Minami (@south37777) 2020年12月28日
上記で言及してない部分で面白い箇所はいっぱいあるので、ぜひ論文をご一読ください!https://t.co/XolZ7fMXFY
— south37/Nao Minami (@south37777) 2020年12月28日
手を動かして学ぶコンテナ標準 - Container Runtime 編
コンテナ標準化の現状と Kubernetes の立ち位置について というブログではコンテナ標準の現状についてまとめてみました。
また、手を動かして学ぶコンテナ標準 - Container Image と Container Registry 編 というブログでは Container Image と Container Registry について手を動かして学んでみました。
このブログでは、runc, containerd などの Container Runtime について、実際に手を動かして学んでみたいと思います。
なお、前回のブログ 同様、基本的にこのブログ内のコマンドは Linux で実行するものとします(自分は MacOS で Vagrant で Ubuntu VM を立てて実験してます)。
runc を動かしてみる
runc は Low-Level Container Runtime と呼ばれるもので、OCI Runtime Specification に準拠した CLI tool となっています。実際に runc を動かしてみることで、Container Runtime に対して理解を深めてみましょう。
まず runc の install ですが、自分が試している ubuntu-20.04 では apt
で install することができます。
vagrant@vagrant:~$ sudo apt update -y vagrant@vagrant:~$ sudo apt install -y runc vagrant@vagrant:~$ which runc /usr/sbin/runc
help を見てみると以下のような内容になっていて、container を操作するために必要な各種 subcommand が存在することがわかります。実は、これらの subcommand によって OCI Runtime Specification の Runtime and Lifecycle で定義された「Operations」 の機能が提供されています。ただし、見比べてみると分かりますが runc 自体はよりリッチな機能を提供しているようです。
vagrant@vagrant:~$ runc --help NAME: runc - Open Container Initiative runtime runc is a command line client for running applications packaged according to the Open Container Initiative (OCI) format and is a compliant implementation of the Open Container Initiative specification. runc integrates well with existing process supervisors to provide a production container runtime environment for applications. It can be used with your existing process monitoring tools and the container will be spawned as a direct child of the process supervisor. Containers are configured using bundles. A bundle for a container is a directory that includes a specification file named "config.json" and a root filesystem. The root filesystem contains the contents of the container. To start a new instance of a container: # runc run [ -b bundle ] <container-id> Where "<container-id>" is your name for the instance of the container that you are starting. The name you provide for the container instance must be unique on your host. Providing the bundle directory using "-b" is optional. The default value for "bundle" is the current directory. USAGE: runc [global options] command [command options] [arguments...] VERSION: spec: 1.0.1-dev COMMANDS: checkpoint checkpoint a running container create create a container delete delete any resources held by the container often used with detached container events display container events such as OOM notifications, cpu, memory, and IO usage statistics exec execute new process inside the container init initialize the namespaces and launch the process (do not call it outside of runc) kill kill sends the specified signal (default: SIGTERM) to the container's init process list lists containers started by runc with the given root pause pause suspends all processes inside the container ps ps displays the processes running inside a container restore restore a container from a previous checkpoint resume resumes all processes that have been previously paused run create and run a container spec create a new specification file start executes the user defined process in a created container state output the state of a container update update container resource constraints help, h Shows a list of commands or help for one command GLOBAL OPTIONS: --debug enable debug output for logging --log value set the log file path where internal debug information is written --log-format value set the format used by logs ('text' (default), or 'json') (default: "text") --root value root directory for storage of container state (this should be located in tmpfs) (default: "/run/user/1000/runc") --criu value path to the criu binary used for checkpoint and restore (default: "criu") --systemd-cgroup enable systemd cgroup support, expects cgroupsPath to be of form "slice:prefix:name" for e.g. "system.slice:runc:434234" --rootless value ignore cgroup permission errors ('true', 'false', or 'auto') (default: "auto") --help, -h show help --version, -v print the version
実際に runc を利用して container を起動してみましょう。runc の README の Using runc というセクションがぴったりの題材なので、これを参考にしてみます。
OCI Runtime Speicification の内容を見てもらうと分かるのですが、runc のような Low-Level Container Runtime は Filesystem Bundle と呼ばれる「Container を起動するのに利用される configuration file と、container にとっての root filesystem となる directory の組み合わせ」を必要とします。まずはこの Filesystem Bundle を用意します。
Filesystem Bundle は config.json と呼ばれる configuration file と、その config.json の中の root.path
で path を指定される 「container の root filesystem となる directory」から構成されます。それぞれ順番に作成していきます。
まずは Filesystem Bundle を格納する directory として mycontainer
という directory を作っておきます。
vagrant@vagrant:~$ mkdir mycontainer vagrant@vagrant:~$ cd mycontainer/
次に、config.json
についてですが、 $ runc spec
を実行すると自動でデフォルト設定の config.json
が作成されます。今回はこれを使うことにしましょう。
vagrant@vagrant:~/mycontainer$ runc spec vagrant@vagrant:~/mycontainer$ ls config.json
デフォルト設定の config.json は以下のような内容になっています。実行時のコマンドや環境変数、root.path などに加えて、namespace や volume mount, capability の設定なども記載されています。
vagrant@vagrant:~/mycontainer$ cat config.json { "ociVersion": "1.0.1-dev", "process": { "terminal": true, "user": { "uid": 0, "gid": 0 }, "args": [ "sh" ], "env": [ "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin", "TERM=xterm" ], "cwd": "/", "capabilities": { "bounding": [ "CAP_AUDIT_WRITE", "CAP_KILL", "CAP_NET_BIND_SERVICE" ], "effective": [ "CAP_AUDIT_WRITE", "CAP_KILL", "CAP_NET_BIND_SERVICE" ], "inheritable": [ "CAP_AUDIT_WRITE", "CAP_KILL", "CAP_NET_BIND_SERVICE" ], "permitted": [ "CAP_AUDIT_WRITE", "CAP_KILL", "CAP_NET_BIND_SERVICE" ], "ambient": [ "CAP_AUDIT_WRITE", "CAP_KILL", "CAP_NET_BIND_SERVICE" ] }, "rlimits": [ { "type": "RLIMIT_NOFILE", "hard": 1024, "soft": 1024 } ], "noNewPrivileges": true }, "root": { "path": "rootfs", "readonly": true }, "hostname": "runc", "mounts": [ { "destination": "/proc", "type": "proc", "source": "proc" }, { "destination": "/dev", "type": "tmpfs", "source": "tmpfs", "options": [ "nosuid", "strictatime", "mode=755", "size=65536k" ] }, { "destination": "/dev/pts", "type": "devpts", "source": "devpts", "options": [ "nosuid", "noexec", "newinstance", "ptmxmode=0666", "mode=0620", "gid=5" ] }, { "destination": "/dev/shm", "type": "tmpfs", "source": "shm", "options": [ "nosuid", "noexec", "nodev", "mode=1777", "size=65536k" ] }, { "destination": "/dev/mqueue", "type": "mqueue", "source": "mqueue", "options": [ "nosuid", "noexec", "nodev" ] }, { "destination": "/sys", "type": "sysfs", "source": "sysfs", "options": [ "nosuid", "noexec", "nodev", "ro" ] }, { "destination": "/sys/fs/cgroup", "type": "cgroup", "source": "cgroup", "options": [ "nosuid", "noexec", "nodev", "relatime", "ro" ] } ], "linux": { "resources": { "devices": [ { "allow": false, "access": "rwm" } ] }, "namespaces": [ { "type": "pid" }, { "type": "network" }, { "type": "ipc" }, { "type": "uts" }, { "type": "mount" } ], "maskedPaths": [ "/proc/acpi", "/proc/asound", "/proc/kcore", "/proc/keys", "/proc/latency_stats", "/proc/timer_list", "/proc/timer_stats", "/proc/sched_debug", "/sys/firmware", "/proc/scsi" ], "readonlyPaths": [ "/proc/bus", "/proc/fs", "/proc/irq", "/proc/sys", "/proc/sysrq-trigger" ] }
次に、container の root filesystem となる directory を用意します。今回は楽をするために、「$ docker export
の機能を利用する」ことにします(自分は vagrant で synced_folder した directory で docker export しました)。原理的には、どんな手段で用意しても良いはずです。
$ mkdir rootfs $ docker export $(docker create busybox) | tar -C rootfs -xvf -
上記コマンドまで実行が終わると、mycontainer
directory の中は以下のような状態になっているはずです。
vagrant@vagrant:~/mycontainer$ ls config.json rootfs
この状態で $ sudo runc run containerid
を実行すると、container が起動します。
vagrant@vagrant:~/oci-playground/runc-playground/mycontainer2$ sudo runc run containerid / #
上記で作成した config.json
はデフォルトで「sh
コマンドを実行する」という設定になっているので、sh
が起動しています。
そのため、一度 exit してから config.json
を例えば「ls
を実行するもの」に書き換えると、ls
が container 内で実行されます。
vagrant@vagrant:~/oci-playground/runc-playground/mycontainer2$ vim config.json # ここで以下のように変更 "args": [ - "sh" + "ls" ], vagrant@vagrant:~/oci-playground/runc-playground/mycontainer2$ sudo runc run containerid bin dev etc home proc root sys tmp usr var
ということで、runc
を利用して無事に container を起動することができました!
config.json + container の root filesystem になる directory という「Filesystem Bundle」を用意さえすれば、「container の起動が簡単に出来る」ということが分かったかと思います。なお、これは OCI Runtime Specification で定められている挙動であるため、この仕様を満たしてさえいれば「runc 以外の Low-Level Container Runtime」においても同様のことができるはずです。
上記の例では runc run
を利用しましたが、その他のオペレーション(create, start, delete など)を利用することでより細かな制御も可能です。詳細が気になる方は、ぜひご自分でも試してみてください。
containerd を動かしてみる
Low-Level Container Runtime の runc の次は、High-Level Container Runtime である containerd を動かしてみましょう。runc は「container を起動する」という部分は実行してくれましたが、それ以外の「Container Image の Pull、Container Image から Filesystem Bundle への変換」などはやってくれませんでした。実は、この部分の機能を提供してくれてるのが containerd です。
containerd は、daemon として常駐して、client からの request を受け付けて動作するという振る舞いになっています。containerd の動作のイメージを掴むために、Getting started というページに従って containerd を利用してみましょう。
まずは、containerd downloads の Installing binaries に従って containerd の binary を download します。そして、その binary を /usr/local/bin
など PATH の通った場所に配置します。
vagrant@vagrant:~$ wget https://github.com/containerd/containerd/releases/download/v1.4.3/containerd-1.4.3-linux-amd64.tar.gz vagrant@vagrant:~$ tar xvf containerd-1.4.3-linux-amd64.tar.gz vagrant@vagrant:~$ sudo mv bin/* /usr/local/bin/
これで、containerd
がコマンドとして利用できるようになります。
vagrant@vagrant:~$ which containerd /usr/local/bin/containerd vagrant@vagrant:~$ containerd --help NAME: containerd - __ _ __ _________ ____ / /_____ _(_)___ ___ _________/ / / ___/ __ \/ __ \/ __/ __ `/ / __ \/ _ \/ ___/ __ / / /__/ /_/ / / / / /_/ /_/ / / / / / __/ / / /_/ / \___/\____/_/ /_/\__/\__,_/_/_/ /_/\___/_/ \__,_/ high performance container runtime USAGE: containerd [global options] command [command options] [arguments...] VERSION: v1.4.3 DESCRIPTION: containerd is a high performance container runtime whose daemon can be started by using this command. If none of the *config*, *publish*, or *help* commands are specified, the default action of the **containerd** command is to start the containerd daemon in the foreground. A default configuration is used if no TOML configuration is specified or located at the default file location. The *containerd config* command can be used to generate the default configuration for containerd. The output of that command can be used and modified as necessary as a custom configuration. COMMANDS: config information on the containerd config publish binary to publish events to containerd oci-hook provides a base for OCI runtime hooks to allow arguments to be injected. help, h Shows a list of commands or help for one command GLOBAL OPTIONS: --config value, -c value path to the configuration file (default: "/etc/containerd/config.toml") --log-level value, -l value set the logging level [trace, debug, info, warn, error, fatal, panic] --address value, -a value address for containerd's GRPC server --root value containerd root directory --state value containerd state directory --help, -h show help --version, -v print the version
ただし、前述したように containerd は daemon として動作する必要があるので、これだけではまだ利用できません。
Getting started を見ると、どうやら containerd のソースコードに同梱されている containerd.service に言及しています。これを利用して、systemd で containerd を起動することにしてみましょう。
vagrant@vagrant:~$ sudo apt install -y unzip # unzip を install しておく vagrant@vagrant:~$ wget https://github.com/containerd/containerd/archive/v1.4.3.zip # ソースコードを取得 vagrant@vagrant:~$ unzip v1.4.3.zip vagrant@vagrant:~$ ls containerd-1.4.3 v1.4.3.zip vagrant@vagrant:~$ ls containerd-1.4.3/containerd.service # systemd 向けの service file が存在 containerd-1.4.3/containerd.service vagrant@vagrant:~$ sudo cp containerd-1.4.3/containerd.service /etc/systemd/system/ vagrant@vagrant:~$ sudo systemctl enable containerd Created symlink /etc/systemd/system/multi-user.target.wants/containerd.service → /etc/systemd/system/containerd.service. vagrant@vagrant:~$ sudo systemctl start containerd
これで、無事 containerd が起動しました。
vagrant@vagrant:~$ ps aux | grep containerd root 13444 1.6 4.5 1344964 45620 ? Ssl 16:39 0:00 /usr/local/bin/containerd vagrant 13455 0.0 0.0 9032 664 pts/0 S+ 16:39 0:00 grep --color=auto containerd
Getting started の document には /etc/containerd/config.toml
の存在も言及されているので、これも用意しておきます。これで準備 OK です。
vagrant@vagrant:~$ sudo mkdir /etc/containerd vagrant@vagrant:~$ sudo vim /etc/containerd/config.toml # ここで containerd の設定ファイルである config.toml を作成 vagrant@vagrant:~$ sudo systemctl restart containerd # ここで、設定ファイルを読み込む形で containerd を再起動
# これが /etc/containerd/config.toml の内容 subreaper = true oom_score = -999 [debug] level = "debug" [metrics] address = "127.0.0.1:1338" [plugins.linux] runtime = "runc" shim_debug = true
さて、containerd が daemon として動くようになりました。次は、containerd と通信するコードを書いてみましょう。Getting started を参考に、以下のような Go コードを書いてみます。
package main import ( "context" "fmt" "log" "syscall" "time" "github.com/containerd/containerd" "github.com/containerd/containerd/cio" "github.com/containerd/containerd/namespaces" "github.com/containerd/containerd/oci" ) func main() { if err := redisExample(); err != nil { log.Fatal(err) } } func redisExample() error { // create a new client connected to the default socket path for containerd client, err := containerd.New("/run/containerd/containerd.sock") if err != nil { return err } defer client.Close() // create a new context with an "example" namespace ctx := namespaces.WithNamespace(context.Background(), "example") // pull the redis image from DockerHub image, err := client.Pull(ctx, "docker.io/library/redis:alpine", containerd.WithPullUnpack) if err != nil { return err } // create a container container, err := client.NewContainer( ctx, "redis-server", containerd.WithImage(image), containerd.WithNewSnapshot("redis-server-snapshot", image), containerd.WithNewSpec(oci.WithImageConfig(image)), ) if err != nil { return err } defer container.Delete(ctx, containerd.WithSnapshotCleanup) // create a task from the container task, err := container.NewTask(ctx, cio.NewCreator(cio.WithStdio)) if err != nil { return err } defer task.Delete(ctx) // make sure we wait before calling start exitStatusC, err := task.Wait(ctx) if err != nil { fmt.Println(err) } // call start on the task to execute the redis server if err := task.Start(ctx); err != nil { return err } // sleep for a lil bit to see the logs time.Sleep(3 * time.Second) // kill the process and get the exit status if err := task.Kill(ctx, syscall.SIGTERM); err != nil { return err } // wait for the process to fully exit and print out the exit status status := <-exitStatusC // Block here. code, _, err := status.Result() if err != nil { return err } fmt.Printf("redis-server exited with status: %d\n", code) // For Debug fmt.Printf("Client: %v\n", client) fmt.Printf("Image: %v\n", image) fmt.Printf("Container: %v\n", container) fmt.Printf("Task: %v\n", task) return nil }
「Go 環境の構築」は良い感じにやってください。その状態で上記の Go コードを実行すると、以下のような出力が行われます。
vagrant@vagrant:~/containerd-playground$ go build main.go && sudo ./main 1:C 10 Dec 2020 17:11:43.852 # oO0OoO0OoO0Oo Redis is starting oO0OoO0OoO0Oo 1:C 10 Dec 2020 17:11:43.852 # Redis version=6.0.9, bits=64, commit=00000000, modified=0, pid=1, just started 1:C 10 Dec 2020 17:11:43.852 # Warning: no config file specified, using the default config. In order to specify a config file use redis-server /path/to/redis.conf 1:M 10 Dec 2020 17:11:43.856 # You requested maxclients of 10000 requiring at least 10032 max file descriptors. 1:M 10 Dec 2020 17:11:43.856 # Server can't set maximum open files to 10032 because of OS error: Operation not permitted. 1:M 10 Dec 2020 17:11:43.856 # Current maximum open files is 1024. maxclients has been reduced to 992 to compensate for low ulimit. If you need higher maxclients increase 'ulimit -n'. 1:M 10 Dec 2020 17:11:43.858 * Running mode=standalone, port=6379. 1:M 10 Dec 2020 17:11:43.858 # Server initialized 1:M 10 Dec 2020 17:11:43.858 # WARNING overcommit_memory is set to 0! Background save may fail under low memory condition. To fix this issue add 'vm.overcommit_memory = 1' to /etc/sysctl.conf and then reboot or run the command 'sysctl vm.overcommit_memory=1' for this to take effect. 1:M 10 Dec 2020 17:11:43.858 * Ready to accept connections 1:signal-handler (1607620306) Received SIGTERM scheduling shutdown... 1:M 10 Dec 2020 17:11:46.877 # User requested shutdown... 1:M 10 Dec 2020 17:11:46.877 * Saving the final RDB snapshot before exiting. 1:M 10 Dec 2020 17:11:46.879 * DB saved on disk 1:M 10 Dec 2020 17:11:46.879 # Redis is now ready to exit, bye bye... redis-server exited with status: 0 Client: &{{<nil> <nil> <nil> <nil> map[] <nil> <nil> <nil> <nil> <nil>} {0 0} 0xc00010ee00 io.containerd.runc.v2 {0xc0002131a0} 0xe6c020} Image: &{0xc0002f4000 {docker.io/library/redis:alpine map[] {application/vnd.docker.distribution.manifest.list.v2+json sha256:b0e84b6b92149194d99953e44f7d1fa1f470a769529bb05b4164eae60d8aea6c 1645 [] map[] <nil>} {312279631 63742833489 <nil>} {749989962 63743217103 <nil>}} {0xc0002131a0}} Container: &{0xc0002f4000 redis-server {redis-server map[] docker.io/library/redis:alpine {io.containerd.runc.v2 <nil>} 0xc0003ce1e0 redis-server-snapshot overlayfs {775955650 63743217103 <nil>} {775955650 63743217103 <nil>} map[]}} Task: &{0xc0002f4000 0xc0001020c0 0xc0003a86c0 redis-server 65235}
上記の Go コードおよび出力について少し説明します。
上記の Go コードは「containerd と /run/containerd/containerd.sock
を通じて通信する client の動作」が記述されています。コード内では client の作成、Container Image の pull、Container Image からの Snapshot (= Container 実行時に root filesystem として利用するもの) および Spec (= Container 実行時のメタデータ、Filesystem Bundle における config.json に相当) の作成、Task (= container の中で実際に実行したいコマンド) の作成、開始、終了およびその待ち合わせを行っています。
また、最後にデバッグ用途に様々な struct を print してみています(これは Getting started に掲載されていたコードに自分が後から追加したものです)。
Pull している Container Image は docker.io/library/redis:alpine
で、ログの出力としても redis server の起動および終了が行われていることが分かります。
これで、containerd を実際に利用してみる事が出来ました!
なお、containerd は OSS でコードが公開されているので、気になった挙動はコードを読んで理解する事が出来ます。例えば、 containerd.WithNewSnapshot
や containerd.WithNewSpec
というのは以下のコードに該当します。
https://github.com/containerd/containerd/blob/v1.4.3/container_opts.go#L144-L169 https://github.com/containerd/containerd/blob/v1.4.3/container_opts.go#L243-L253
前者のコードを読み進めると「Snapshotter によって Snapshot 作成を行って、Container には SnapshotKey を設定して id を通して参照できるようにしている」事が分かりますし、後者のコードを読み進めると「OCI Runtime Specification の Go binding で定義された Spec
struct が generate されて利用される」事が分かります。「runc などの Low-Level Container Runtime が必要とする情報を用意する」という部分をしっかりやってくれてるようです(注: これは自分がコードを読んだ理解なので、誤解してる可能もあります。何か気がついた際はコメントいただけるとありがたいです)。
その他、Container Image を Pull している部分や Task の管理部分なども興味深い部分です。気になった箇所はぜひ読み進めて理解を深めてみてください。
まとめ
runc や containerd などの Container Runtime を実際に触ってみる事で、理解を深めました。runc については、OCI Runtime Specification で定義された挙動が CLI tool として提供されている事が分かりました。containerd については、Container Image の Pull や Container Image から Filesystem Bundle 相当の情報(= Container 実行に必要な Snapshot および Spec)への変換、Container や Task の作成・削除・開始などの管理を行っている事が分かりました。
runc やcontainerd は扱っている領域が異なる一方で、どちらも「Container Runtime」という名前で呼ばれるために混乱してしまいがちです。実際に手を動かしてみる事で、「それぞれの責任範囲」や「レイヤー構造」について理解が深まったように思います。
なお、このブログ自体は自分の理解のために試したことをまとめたものですが、誰か他の人にとっても理解を助けるものになっていればとても幸いです。
Vagrant での実験環境
以下のような Vagrantfile を使ってます。ubuntu-20.04 を使ってます。
# -*- mode: ruby -*- # vi: set ft=ruby : VAGRANTFILE_API_VERSION = "2" Vagrant.configure(VAGRANTFILE_API_VERSION) do |config| config.vm.box = "bento/ubuntu-20.04" config.vm.synced_folder "./", "/home/vagrant/oci-playground" end
手を動かして学ぶコンテナ標準 - Container Image と Container Registry 編
先日は、コンテナ標準化の現状と Kubernetes の立ち位置について において、各種ドキュメントをベースにコンテナ標準についてまとめてみました。
このブログでは、実際に tool などに触れて手を動かすことで、コンテナ標準についてさらに理解を深めてみたいと思います。
なお、基本的にこのブログ内のコマンドは、Linux で実行するものとします(自分は MacOS で Vagrant で Ubuntu VM を立てて実験してます)。
OCI Image の中身を見てみる
skopeo と呼ばれる「container image に対して様々な操作を行えるツール」があります。このツールを利用することで、「docker image から OCI Image への変換」を行うことができます。このツールを利用して、実際に OCI Image の中身を見てみましょう。
まず、以下のコマンドを実行して ruby:2.7.2-slim
という docker image を oci:ruby-oci:2.7.2
という名前の OCI Image に変換します。
vagrant@vagrant:~/oci-playground$ skopeo copy docker://ruby:2.7.2-slim oci:ruby-oci:2.7.2 Getting image source signatures Copying blob 852e50cd189d done Copying blob 6de4319615e2 done Copying blob 150eb06190d1 done Copying blob cf654ff9d9df done Copying blob 0a529f6cf42e done Copying config 3265430f5e done Writing manifest to image destination Storing signatures
上記コマンドを実行すると、ruby-oci という directory が出来ています。
vagrant@vagrant:~/oci-playground$ ls ruby-oci
ruby-oci
directory の中を見てみると、以下のように blobs
という direcyory と index.json
, oci-layout
という file が出来ています。これは、OCI Image Format Specification で定められた Image Layout の内容に一致しています。
vagrant@vagrant:~/oci-playground/ruby-oci$ ls blobs index.json oci-layout
oci-layout file には imageLayoutVersion
だけが記載されています。現時点では 1.0.0
が記載されているだけなので、将来の拡張のための file と考えると良いでしょう。
vagrant@vagrant:~/oci-playground/ruby-oci$ cat oci-layout | jq . { "imageLayoutVersion": "1.0.0" }
index.json は OCI Image のエントリーポイントとも呼べる file で、ここには以下のように「manifest fileへの参照(= Image Manifest を指し示す Content Descriptor)」が記載されています。
vagrant@vagrant:~/oci-playground/ruby-oci$ cat index.json | jq . { "schemaVersion": 2, "manifests": [ { "mediaType": "application/vnd.oci.image.manifest.v1+json", "digest": "sha256:ad39959791540e6213fbe4675b9f3ee11e96456df3601b0936973ca7ae766bd7", "size": 976, "annotations": { "org.opencontainers.image.ref.name": "2.7.2" } } ] }
ここで出てきた「Content Descriptor」というのが OCI Image Format において特徴的なもので、これは「mediaType
, digest
, size
の 3 つ組 + optional な情報 (e.g. annotations)」となっています。
mediaType
が参照先の情報の種類、digest
が参照先の情報の path、size
が参照先の情報のバイト数を表しています。
digest で示されているのは「blobs
directory 以下の file path」になっていて、例えば上記の sha256:ad39959791540e6213fbe4675b9f3ee11e96456df3601b0936973ca7ae766bd7
という digest は blobs/sha256/ad39959791540e6213fbe4675b9f3ee11e96456df3601b0936973ca7ae766bd7
という path を表しています。実際に、file の中身を見てみると以下のような JSON になっています。
vagrant@vagrant:~/oci-playground/ruby-oci$ cat blobs/sha256/ad39959791540e6213fbe4675b9f3ee11e96456df3601b0936973ca7ae766bd7 | jq . { "schemaVersion": 2, "config": { "mediaType": "application/vnd.oci.image.config.v1+json", "digest": "sha256:3265430f5e5babe0664d7f7bcc77db2ef7d5feaa1625c06c10b1409ad2952133", "size": 4598 }, "layers": [ { "mediaType": "application/vnd.oci.image.layer.v1.tar+gzip", "digest": "sha256:852e50cd189dfeb54d97680d9fa6bed21a6d7d18cfb56d6abfe2de9d7f173795", "size": 27105484 }, { "mediaType": "application/vnd.oci.image.layer.v1.tar+gzip", "digest": "sha256:6de4319615e27e1aaaadc89b43db39ea0e118f47eeecfa4c8b910ca2fd810653", "size": 12539406 }, { "mediaType": "application/vnd.oci.image.layer.v1.tar+gzip", "digest": "sha256:150eb06190d1ba56f7b998da25a140c21258bca436d33e2e77df679d77ab364a", "size": 198 }, { "mediaType": "application/vnd.oci.image.layer.v1.tar+gzip", "digest": "sha256:cf654ff9d9df475122683b6bd070fa57a1e1969ced2a45f2c1f76a0678495ef2", "size": 22852677 }, { "mediaType": "application/vnd.oci.image.layer.v1.tar+gzip", "digest": "sha256:0a529f6cf42e0fb49fe3fb4d12e232b26db923ab85a442563b0a7ae0a28c5971", "size": 143 } ] }
mediaType
が application/vnd.oci.image.manifest.v1+json
だったことから、これは Image Manifest であると分かります。実際に、Image Manifest の仕様で定義された内容と一致しており、config (Container Image のメタデータ)や layers (Container Image の Layer、Docker Image における Layer Cache の単位となるもの)を情報として持つことも分かります。また、それらの情報への参照も、先ほどと同様の Content Descriptor 形式で表されていることが分かります。
config
の内容は、以下のような Image Configuration となっています。環境変数や Command など Container 実行時に必要な各種メタデータや、Container Image 作成時の history の情報が記載されています。
vagrant@vagrant:~/oci-playground/ruby-oci$ cat blobs/sha256/3265430f5e5babe0664d7f7bcc77db2ef7d5feaa1625c06c10b1409ad2952133 | jq . { "created": "2020-11-18T15:35:15.373100656Z", "architecture": "amd64", "os": "linux", "config": { "Env": [ "PATH=/usr/local/bundle/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin", "LANG=C.UTF-8", "RUBY_MAJOR=2.7", "RUBY_VERSION=2.7.2", "RUBY_DOWNLOAD_SHA256=1b95ab193cc8f5b5e59d2686cb3d5dcf1ddf2a86cb6950e0b4bdaae5040ec0d6", "GEM_HOME=/usr/local/bundle", "BUNDLE_SILENCE_ROOT_WARNING=1", "BUNDLE_APP_CONFIG=/usr/local/bundle" ], "Cmd": [ "irb" ] }, "rootfs": { "type": "layers", "diff_ids": [ "sha256:f5600c6330da7bb112776ba067a32a9c20842d6ecc8ee3289f1a713b644092f8", "sha256:70ca8ae918406dce7acc5fe0f49e45b9275a266b83e275922e67358976c2929e", "sha256:e8ace463e6f7085a5439cf3b578a080fbefc8ad8424b59b9f35590adb1509763", "sha256:71e4ad27368acf7dbb5c90aa65d67cc462267836aa220cbafb9bb62acd9d48de", "sha256:1946ed62a3cb062940077a7a1dbfc93d55be6ef3d4f605883b42f71970381662" ] }, "history": [ { "created": "2020-11-17T20:21:17.570073346Z", "created_by": "/bin/sh -c #(nop) ADD file:d2abb0e4e7ac1773741f51f57d3a0b8ffc7907348842d773f8c341ba17f856d5 in / " }, { "created": "2020-11-17T20:21:17.865210281Z", "created_by": "/bin/sh -c #(nop) CMD [\"bash\"]", "empty_layer": true }, { "created": "2020-11-18T15:21:22.717162717Z", "created_by": "/bin/sh -c set -eux; \tapt-get update; \tapt-get install -y --no-install-recommends \t\tbzip2 \t\tca-certificates \t\tlibffi-dev \t\tlibgmp-dev \t\tlibssl-dev \t\tlibyaml-dev \t\tprocps \t\tzlib1g-dev \t; \trm -rf /var/lib/apt/lists/*" }, { "created": "2020-11-18T15:21:23.811888513Z", "created_by": "/bin/sh -c set -eux; \tmkdir -p /usr/local/etc; \t{ \t\techo 'install: --no-document'; \t\techo 'update: --no-document'; \t} >> /usr/local/etc/gemrc" }, { "created": "2020-11-18T15:21:24.004412503Z", "created_by": "/bin/sh -c #(nop) ENV LANG=C.UTF-8", "empty_layer": true }, { "created": "2020-11-18T15:30:41.383881949Z", "created_by": "/bin/sh -c #(nop) ENV RUBY_MAJOR=2.7", "empty_layer": true }, { "created": "2020-11-18T15:30:41.629378277Z", "created_by": "/bin/sh -c #(nop) ENV RUBY_VERSION=2.7.2", "empty_layer": true }, { "created": "2020-11-18T15:30:41.868222399Z", "created_by": "/bin/sh -c #(nop) ENV RUBY_DOWNLOAD_SHA256=1b95ab193cc8f5b5e59d2686cb3d5dcf1ddf2a86cb6950e0b4bdaae5040ec0d6", "empty_layer": true }, { "created": "2020-11-18T15:35:11.770005784Z", "created_by": "/bin/sh -c set -eux; \t\tsavedAptMark=\"$(apt-mark showmanual)\"; \tapt-get update; \tapt-get install -y --no-install-recommends \t\tautoconf \t\tbison \t\tdpkg-dev \t\tgcc \t\tlibbz2-dev \t\tlibgdbm-compat-dev \t\tlibgdbm-dev \t\tlibglib2.0-dev \t\tlibncurses-dev \t\tlibreadline-dev \t\tlibxml2-dev \t\tlibxslt-dev \t\tmake \t\truby \t\twget \t\txz-utils \t; \trm -rf /var/lib/apt/lists/*; \t\twget -O ruby.tar.xz \"https://cache.ruby-lang.org/pub/ruby/${RUBY_MAJOR%-rc}/ruby-$RUBY_VERSION.tar.xz\"; \techo \"$RUBY_DOWNLOAD_SHA256 *ruby.tar.xz\" | sha256sum --check --strict; \t\tmkdir -p /usr/src/ruby; \ttar -xJf ruby.tar.xz -C /usr/src/ruby --strip-components=1; \trm ruby.tar.xz; \t\tcd /usr/src/ruby; \t\t{ \t\techo '#define ENABLE_PATH_CHECK 0'; \t\techo; \t\tcat file.c; \t} > file.c.new; \tmv file.c.new file.c; \t\tautoconf; \tgnuArch=\"$(dpkg-architecture --query DEB_BUILD_GNU_TYPE)\"; \t./configure \t\t--build=\"$gnuArch\" \t\t--disable-install-doc \t\t--enable-shared \t; \tmake -j \"$(nproc)\"; \tmake install; \t\tapt-mark auto '.*' > /dev/null; \tapt-mark manual $savedAptMark > /dev/null; \tfind /usr/local -type f -executable -not \\( -name '*tkinter*' \\) -exec ldd '{}' ';' \t\t| awk '/=>/ { print $(NF-1) }' \t\t| sort -u \t\t| xargs -r dpkg-query --search \t\t| cut -d: -f1 \t\t| sort -u \t\t| xargs -r apt-mark manual \t; \tapt-get purge -y --auto-remove -o APT::AutoRemove::RecommendsImportant=false; \t\tcd /; \trm -r /usr/src/ruby; \t! dpkg -l | grep -i ruby; \t[ \"$(command -v ruby)\" = '/usr/local/bin/ruby' ]; \truby --version; \tgem --version; \tbundle --version" }, { "created": "2020-11-18T15:35:12.227711802Z", "created_by": "/bin/sh -c #(nop) ENV GEM_HOME=/usr/local/bundle", "empty_layer": true }, { "created": "2020-11-18T15:35:12.563337139Z", "created_by": "/bin/sh -c #(nop) ENV BUNDLE_SILENCE_ROOT_WARNING=1 BUNDLE_APP_CONFIG=/usr/local/bundle", "empty_layer": true }, { "created": "2020-11-18T15:35:12.907595531Z", "created_by": "/bin/sh -c #(nop) ENV PATH=/usr/local/bundle/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin", "empty_layer": true }, { "created": "2020-11-18T15:35:14.977063521Z", "created_by": "/bin/sh -c mkdir -p \"$GEM_HOME\" && chmod 777 \"$GEM_HOME\"" }, { "created": "2020-11-18T15:35:15.373100656Z", "created_by": "/bin/sh -c #(nop) CMD [\"irb\"]", "empty_layer": true } ] }
layers
は Filesystem Layer を表しています。tar+gzip
という mediaType の suffix は「gzip 圧縮された tar archive」を表しています。試しに、最も root にあった sha256:852e50cd189dfeb54d97680d9fa6bed21a6d7d18cfb56d6abfe2de9d7f173795
の中身を見てみます。
vagrant@vagrant:~/oci-playground/ruby-oci$ mkdir rootfs vagrant@vagrant:~/oci-playground/ruby-oci$ tar xvzf blobs/sha256/852e50cd189dfeb54d97680d9fa6bed21a6d7d18cfb56d6abfe2de9d7f173795 -C rootfs/ . . .
上記コマンドで、rootfs directory 以下に圧縮されていた中身が展開されます(注: tar: bin/uncompress: Cannot hard link to ‘bin/gunzip’: Operation not permitted
など一部の file について error は出ていて、そのせいで tar: Exiting with failure status due to previous errors
という失敗 message も出てしまいましたが、それはここでは無視します)。
rootfs
の中身を見てみると、以下のようにいくつかの directyory が並んでいます。
vagrant@vagrant:~/oci-playground/ruby-oci$ ls rootfs/ bin boot dev etc home lib lib64 media mnt opt proc root run sbin srv sys tmp usr var
試しに変換前の ruby:2.7.2-slim
docker image を利用して container を起動してみると、root directory の中身がそっくりであることが確認できます。
$ docker run -it ruby:2.7.2-slim bash root@f6be3c7c619d:/# ls / bin boot dev etc home lib lib64 media mnt opt proc root run sbin srv sys tmp usr var
実は、これらの layer が「どう apply されるのか」は Image Layer Filesystem Changeset の Applying Changesets というセクション の中で以下のように明確に定義されています。ざっくり言えば「layer の上から順に tar archive を展開したようなもの」になります。 「file や directory の削除」は Whiteouts と呼ばれる特別な仕様で表現されますが、特別に注意を払う必要があるのはそれくらいのようです。
Applying Changesets
- Layer Changesets of media type application/vnd.oci.image.layer.v1.tar are applied, rather than simply extracted as tar archives.
- Applying a layer changeset requires special consideration for the whiteout files.
- In the absence of any whiteout files in a layer changeset, the archive is extracted like a regular tar archive.
cf. https://github.com/opencontainers/image-spec/blob/v1.0.1/layer.md#applying-changesets
ということで、OCI image の中身に目を通してみました。「Conatainer を走らせるために必要な情報(= 実行時のメタデータ + Layer 化された filesystem の情報)」が格納されてることがわかったかと思います。
Container Registry との通信内容を見てみる
ここまでで、「Container Image の内容」については把握できました。次に、「Container Registry から Container Image をどのように pull しているのか」を調べてみましょう。
現在、各種 Container Registry は Docker 社が公開している Docker Registry HTTP API V2 と呼ばれる仕様に従う形で Container Image の Pull を出来るようにしています。実は、「Container Image の Pull」にあたる操作はただの HTTP request であるため、$ curl
を利用して実行する事ができます。ここでは、実際に $ curl
で request してみることで、Container Registry との通信内容を見てみる事にしましょう。
なお、自分が試した範囲では、どの Container Image も OCI Image Format ではなく Docker Image Manifest V 2, Schema 2 に従う形の response を返してきました。ただ、OCI Image Format と Docker Image V2.2 は一部の mediaType 名を除いてほぼ同一なので、先ほど眺めた内容は理解に役立つはずです。
さて、実際に curl で request を送ってみましょう。対象 Container Image は何でも良いのですが、ここでは https://github.com/GoogleContainerTools/base-images-docker に記載されてる Debian の Container Image である gcr.io/google-appengine/debian9
を対象にしてみます。
まず、以下のように Container Registry の Authentication に必要な Token を取得します。この時、「google-appengine/debian9
の pull」という形で scope
を指定しておきます。
$ export TOKEN=$(curl "https://gcr.io/v2/token?service=gcr.io&scope=repository:google-appengine/debian9:pull" | jq -r '.token') % Total % Received % Xferd Average Speed Time Time Time Current Dload Upload Total Spent Left Speed 100 453 0 453 0 0 1088 0 --:--:-- --:--:-- --:--:-- 1088
次に、https://gcr.io/v2/<name>/manifests/<reference>
へ先ほど取得した Token 付きで GET
request を送ります。こうすると、Docker Image V2.2 における manifest file が取得できます。
$ curl -H "Authorization: Bearer ${TOKEN}" https://gcr.io/v2/google-appengine/debian9/manifests/latest | jq . % Total % Received % Xferd Average Speed Time Time Time Current Dload Upload Total Spent Left Speed 100 528 100 528 0 0 469 0 0:00:01 0:00:01 --:--:-- 469 { "schemaVersion": 2, "mediaType": "application/vnd.docker.distribution.manifest.v2+json", "config": { "mediaType": "application/vnd.docker.container.image.v1+json", "size": 463, "digest": "sha256:18c47921b263ac67af3d654e3b485c998d1e6bab56edc5a15b6b7a8fad3ac18a" }, "layers": [ { "mediaType": "application/vnd.docker.image.rootfs.diff.tar.gzip", "size": 47965538, "digest": "sha256:faa9d9046d25e5fd30ac4444c7b6c30a1a6fff7c461410547156aed2001668a1" } ] }
まず、config
の中身を見てみましょう。
digest を利用して参照を辿る際は https://gcr.io/v2/<name>/blobs/<digest>
へ request すれば良いです。実際に request してみると、以下のような response が返ってきます。先ほど OCI Image の中身を見てみた時と同様に、Container 実行に必要なメタデータが格納されていることが分かります。
$ curl -L -H "Authorization: Bearer ${TOKEN}" https://gcr.io/v2/google-appengine/debian9/blobs/sha256:18c47921b263ac67af3d654e3b485c998d1e6bab56edc5a15b6b7a8fad3ac18a | jq . % Total % Received % Xferd Average Speed Time Time Time Current Dload Upload Total Spent Left Speed 100 13 0 13 0 0 23 0 --:--:-- --:--:-- --:--:-- 23 100 463 100 463 0 0 750 0 --:--:-- --:--:-- --:--:-- 750 { "architecture": "amd64", "author": "Bazel", "created": "1970-01-01T00:00:00Z", "history": [ { "author": "Bazel", "created": "1970-01-01T00:00:00Z", "created_by": "bazel build ..." } ], "os": "linux", "rootfs": { "type": "layers", "diff_ids": [ "sha256:0a3dcb016bd8a852985044291de00ad6a6b94dcb0eac01b34b56afed409b9999" ] }, "config": { "Cmd": [ "/bin/sh", "-c", "/bin/bash" ], "Env": [ "DEBIAN_FRONTEND=noninteractive", "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin", "PORT=8080" ] } }
なお、注意点として、どうやら GCR は https://gcr.io/v2/<name>/blobs/<digest>
への request では Google Cloud Storage への redirect response を返すようです。-L
オプションを付けない場合は以下のような結果になることには留意してください。
$ curl --include -H "Authorization: Bearer ${TOKEN}" https://gcr.io/v2/google-appengine/debian9/blobs/sha256:18c47921b263ac67af3d654e3b485c998d1e6bab56edc5a15b6b7a8fad3ac18a HTTP/2 302 docker-distribution-api-version: registry/2.0 location: https://storage.googleapis.com/artifacts.google-appengine.appspot.com/containers/images/sha256:18c47921b263ac67af3d654e3b485c998d1e6bab56edc5a15b6b7a8fad3ac18a content-type: application/json date: Wed, 09 Dec 2020 13:38:51 GMT server: Docker Registry cache-control: private x-xss-protection: 0 x-frame-options: SAMEORIGIN alt-svc: h3-29=":443"; ma=2592000,h3-T051=":443"; ma=2592000,h3-Q050=":443"; ma=2592000,h3-Q046=":443"; ma=2592000,h3-Q043=":443"; ma=2592000,quic=":443"; ma=2592000; v="46,43" accept-ranges: none vary: Accept-Encoding {"errors":[]}%
上記では config の取得を行いましたが、Layer (mediaType: application/vnd.docker.image.rootfs.diff.tar.gzip
のデータ)についても同様にhttps://gcr.io/v2/<name>/blobs/<digest>
への request によって取得する事ができます。先ほどと同様に、tar
コマンドで展開すると container 実行に利用される file を取得することが出来ます。
$ curl -L -H "Authorization: Bearer ${TOKEN}" https://gcr.io/v2/google-appengine/debian9/blobs/sha256:faa9d9046d25e5fd30ac4444c7b6c30a1a6fff7c461410547156aed2001668a1 --output /tmp/faa9d9046d25e5fd30ac4444c7b6c30a1a6fff7c461410547156aed2001668a1 % Total % Received % Xferd Average Speed Time Time Time Current Dload Upload Total Spent Left Speed 100 13 0 13 0 0 28 0 --:--:-- --:--:-- --:--:-- 28 100 45.7M 100 45.7M 0 0 13.6M 0 0:00:03 0:00:03 --:--:-- 20.1M $ ls -la /tmp/faa9d9046d25e5fd30ac4444c7b6c30a1a6fff7c461410547156aed2001668a1 -rw-r--r-- 1 minami wheel 47965538 Dec 9 23:48 /tmp/faa9d9046d25e5fd30ac4444c7b6c30a1a6fff7c461410547156aed2001668a1 $ mkdir /tmp/rootfs $ tar xvzf /tmp/faa9d9046d25e5fd30ac4444c7b6c30a1a6fff7c461410547156aed2001668a1 -C /tmp/rootfs/ $ ls /tmp/rootfs bin boot dev etc home lib lib64 media mnt opt proc root run sbin srv sys tmp usr var
という事で、Container Registry との通信について、特に「Container Image の Pull」に絞って通信内容を見てみました。Docker Image V2.2 をベースにした通信である事、特に config や layer などがそれぞれの単位で通信できることなどが分かったかと思います。より詳しい内容が気になる場合は、Docker Registry HTTP API V2 を参照してみてください。
なお、「Container Image 全てをまとめた file を一括でダウンロードしないのは何故なのか?」という疑問についてですが、これは自分の理解では「Layer Cache を効かせた形での Image Pull を実現するため」だと捉えています。Layer のデータは巨大であるため最低限の通信で済ませたいというのが大前提にあり、そのために「コンテンツの中身を反映した digest 値を用いて、Layer ごとに通信する」という振る舞いになっているのだと思われます。
まとめ
OCI image を実際に作成して眺めて見ることで、Container Image について理解を深めました。また、curl
で Container Registry との通信を行うことで、Container Registry との通信内容についても理解を深める事が出来ました。
ドキュメントを読むだけだとどうしても理解が曖昧になってしまいがちですが、実際に手を動かす事で具体的な動作をイメージ出来るようになります。このブログ自体は自分の理解のために試したことをまとめたものですが、誰か他の人にとっても理解を助けるものになっていれば幸いです。
なお、今回は Container Image + Container Registry 編でしたが、後日 Container Runtime についても「手を動かして調べた内容」についてまとめたいと思っています。特に、「Container Image から Container Runtime が利用する Filesystem Bundle への Conversion」や、「runc などの low-level Container Runtime の動作」、「containerd や CRI-O などの high-level Container Runtime の動作」について試したことをまとめる予定です。
補足: Vagrant での実験環境
以下のような Vagrantfile を使ってます。ubuntu-20.04 を使ってます。
# -*- mode: ruby -*- # vi: set ft=ruby : VAGRANTFILE_API_VERSION = "2" Vagrant.configure(VAGRANTFILE_API_VERSION) do |config| config.vm.box = "bento/ubuntu-20.04" end
参考文献
以下のブログは、ツールやコマンド、内容において大幅に参考にさせて頂きました。ありがとうございました。
コンテナ標準化の現状と Kubernetes との関係性について
コンテナ標準化が進んでいる事は知りつつも、標準化された仕様の具体的な内容についてはあまり知らない事に気づいたので、この機会に調べてみました。個人向けメモとして残しておきます。
余力があれば、後でもう少し詳細をまとめる予定です(docker image を OCI Image Format に変換して眺めてみたり、runc や containerd などを実際に動かしてみたり、containerd や CRI-O などの Container Runtime の実装に目を通してみたりしたので、その辺りについてもいつかまとめたいと思ってます)。
追記: 手を動かして調べた内容は以下の2つのブログにまとめました。
以下、調査した内容をまとめたメモです。
コンテナ標準と Open Container Initiative (OCI) について
コンテナ標準は Open Container Initiative (OCI) と呼ばれる団体によって仕様策定が進められている。 image format, (low-level) runtime については既に標準が存在している(v1.0.0 をリリース済み)。distribution (container registry 周り) については仕様策定中の状態(2020年12月6日時点で GitHub の tag では v1.0.0-rc1 が出ている状態)。
以下、 OCI が定める各種標準仕様について簡単にまとめる。
OCI Image Format Specification
- Docker Image のような「container を記述する image format」の標準仕様。これは Docker Image の最新 format である Docker Image Manifest V2, Schema 2 をベースに標準化したもの。
- 以下が Docker 社からの公式声明。
Given this state of the world in late 2015, the OCI image specification work began in earnest with a strong group of collaborating independent and vendor-associated participants, using the Docker v2.2 image format as a starting point.
- cf. https://opencontainers.org/posts/blog/2018-10-11-oci-image-support-comes-to-open-source-docker-registry/
- Container Registry とのデータのやり取りも OCI Image Format(およびそれの元になった Docker Image Manifest V2, Schema2)に準拠する形で行われる。
- 後述する OCI Runtime Specification で利用される Filesystem Bundle を生成する Conversion 処理 についても仕様が策定されている。
OCI Runtime Specification
- Container 管理を行う Container Runtime の標準仕様。後述するが、Low-Level Container Runtime と呼ばれるものはこの仕様に準拠している。
- これは Docker 社が OCI に寄贈した "runC" とほぼ対応づくもの。
runc depends on and tracks the runtime-spec repository. We will try to make sure that runc and the OCI specification major versions stay in lockstep. This means that runc 1.0.0 should implement the 1.0 version of the specification.
- https://github.com/opencontainers/runc
- 「コンテナの configuration file およびコンテナの root filesystem をまとめたもの」である Filesystem Bundle や、OCI Runtime 準拠の Container Runtime で行える操作 について仕様を策定している。
OCI Distribution Specification
- Container Registry との通信における標準仕様。Docker Hub や Google Container Registry など各種 Container Registry が従っている Docker Registry HTTP API V2 をベースに仕様策定が進められている。最新は v1.0.0-rc1。
The spec is based on the specification for the Docker Registry HTTP API V2 protocol apdx-1.
- cf. https://github.com/opencontainers/distribution-spec/blob/master/spec.md
Container Runtime について
Container Runtime と呼ばれるものは複数存在するが、ものによって担当する layer が違う。以下、一例を紹介(注: 他にも Container Runtime と呼ばれるものはいくつかあるが、ここでは自分の理解のために調べた一部に限定している)。
runc
- 前述した OCI Runtime Specification を素朴に実装したもの。後述する containerd と対比して、Low-Level Container Runtime と紹介をされる例を見かける。
runc is a low-level container runtime, as it directly uses namespace and cgroups to create containers.
- https://insujang.github.io/2019-10-31/container-runtime
- Docker 社が「ツールとしての Docker」の一部を OSS として公開したのが出自。
And today we are spinning out runC as a standalone tool, to be used as plumbing by infrastructure plumbers everywhere.
- https://www.docker.com/blog/runc/
- OCI Runtime Specification で定められた Filesystem Bundle と呼ばれる「コンテナの configuration file およびコンテナの root filesystem をまとめたもの」を元に、コンテナの作成、削除、状態の取得などの操作が可能。
- README の Using runc を参照して動かしてみると動作がイメージできる。
containerd
- container registry からの image 取得、runc などの low-level container runtime を利用した container 起動(OCI Runtime Specification で定められた filesystem bundle 生成を含む)、container 管理などを行うもの。
- daemon として動作する。client からは
/run/containerd/containerd.sock
経由の gRPC API で通信を行い、container 作成や task 実行などの操作が可能。 - 後述する Container Runtime Interface (CRI) もサポートしている
- 「ツールとしての Docker」から利用されている。そもそも、Docker 社が CNCF へ寄贈したのが出自。
Today, Docker announced its intention to donate the containerd project to the Cloud Native Computing Foundation (CNCF).
- https://www.docker.com/blog/docker-donates-containerd-to-cncf/
CRI-O
- containerd とだいたい似たレイヤーを担当。「Kubernetes から利用されること(CRI に準拠していること)」を念頭に開発されてる。
Last year, the Kubernetes project introduced its Container Runtime Interface (CRI)
Building on that work, the CRI-O project (originally known as OCID) is ready to provide a lightweight runtime for Kubernetes.
- https://www.redhat.com/en/blog/introducing-cri-o-10
- runc を始めとして、いくつかのコンポーネントを組み合わせた実装になっている
The plan is to use OCI projects and best of breed libraries for different aspects:
- Runtime: runc (or any OCI runtime-spec implementation) and oci runtime tools
- Images: Image management using containers/image
- Storage: Storage and management of image layers using containers/storage
- Networking: Networking support through use of CNI
- https://github.com/cri-o/cri-o#what-is-not-in-scope-for-this-project
Kubernetes が定める Container Runtime Interface (CRI) について
- Kubernetes は Container Runtime Interface (CRI) と呼ばれる「独自で定義した API」を利用して、Container Runtime と通信を行う
- CRI は基本的には「protocol buffer で記述された gRPC API」。container を操作するために必要な操作が RPC として定義されている。
CRI consists of a protocol buffers and gRPC API, and libraries, with additional specifications and tools under active development
- cf. https://kubernetes.io/blog/2016/12/container-runtime-interface-cri-in-kubernetes/
- アクティブに開発されてる Container Runtime は、CRI を実装している。一例は以下(注: Docker については Kubernetes 開発者によって dockershim と呼ばれる CRI サポート用のツールが実装されていた)。
This page lists details for using several common container runtimes with Kubernetes, on Linux:
- containerd
- CRI-O
- Docker
- https://kubernetes.io/docs/setup/production-environment/container-runtimes/
- 先週、 Docker が Kubernetes の Container Runtime としては Deprecated になった ことが話題になったが、これは dockershim のメンテナンスを将来のバージョンで止めるというアナウンス。
- 自分が調べた限りでは、containerd, CRI-O などの Container Runtime はどれも複数の image format をサポートしていて、「docker build で生成した container image」は引き続き利用可能。
- "Docker Image Manifest V2, Schema 2" と "OCI Image Format Specification" はどちらもサポートされている。
- そもそもこれらは「ほぼ同一」ある(一部の mediaType が違うくらいで、1対1対応)
The OCI image manifest is also a registry image manifest that defines components that make up an image. The format is essentially the same as the Docker V2.2 format, with a few differences.
- mediaType - must be set to application/vnd.oci.image.manifest.v1+json
- config.mediaType - must be set to application/vnd.oci.image.config.v1+json
Each object in layers must have mediaType be either application/vnd.oci.image.layer.v1.tar+gzip or application/vnd.oci.image.layer.v1.tar.
- https://containers.gitbook.io/build-containers-the-hard-way/#registry-format-oci-image-manifest
- そもそもこれらは「ほぼ同一」ある(一部の mediaType が違うくらいで、1対1対応)
- "Docker Image Manifest V2, Schema 2" と "OCI Image Format Specification" はどちらもサポートされている。
- そのため、「docker build で作成した container image を container registry へ push して、k8s から利用する」という一連のワークフローに関していえば、containerd or CRI-O を使っても問題になる事は一切無い はず
- もちろん、以下の Blog で言及されてるように「Kuberrnetes 内で Docker 自体の機能に依存していた場合」は対応が必要
2017、2018 年の振り返り
年が明けて 2019 年になりました。良い機会なので、2017、2018 年の2年間を振り返ってみたいと思います。
この2年間は色々な変化がありました。仕事面とプライベート面の2つに大きく分けてまとめてみます。
仕事面の振り返り
2017 年は、スカウトというプロダクトのグロースを主に行なっていました。弊社(Wantedly)ではプロダクトごとにチームがあり、自分はチームの中の Tech Lead & Product Manager として、「アイディアを機能に落とし込み、プロダクトを成長させる」という役割を担っていました。ポジションやチームは変わりつつも、2016年以前とはそれほど変わらない働き方でした。
一方、2018 年はインフラチームへ移籍をしたため、働き方が大きく変わりました。
弊社のインフラチームは、「SRE」と「技術基盤」という2つの役割を持っています。SRE としては、「安定したサービスの提供」に責任を持っており、SLI/SLA の策定や監視、障害対応、サービスの可用性やスケーラビリティを向上させる取り組みなどを行なっています。技術基盤としては、「エンジニアの生産性を最大化させること」に責任を持っており、様々な CLI ツールの作成や機械学習基盤の構築、CI/CD フローの構築・改善などを行なっています。また、弊社のマイクロサービスアーキテクチャをより良いものへ作り変えることや、社内横断的に必要なマイクロサービス(e.g. ユーザーアカウントを扱うサービス、ログインフローを扱うサービスなど)の開発も技術基盤として取り組んでいます。
インフラチームへ移ったことで、より社内横断的に必要な開発を行うようになりました。
移籍後にやった事は、細かいものを除くと以下の2つです。
- 機械学習基盤の整備
- 社内横断的に必要なマイクロサービスの開発
機械学習基盤の整備
機械学習基盤については、「社内の機械学習エンジニアの生産性を高める」事をゴールにプロジェクトを進めました。成果は Rejectcon や AWS Dev Day で発表しています。会社としても個人としても、やった事をちゃんと外に出せたのは良かったかなと思います。
https://speakerdeck.com/south37/number-rejectcon2018 https://speakerdeck.com/south37/number-awsdevday
また、「機械学習の評価データ管理」のためにいくつかツールを作っていて、それらは OSS として公開しました。
後で「プライベート面の振り返り」で詳しく書きますが、2017年頃から自分自身が OSS 活動に積極的になり、仕事でもそれが活かされてきたかなと思います。
社内横断的に必要なマイクロサービスの開発
社内横断的に必要なマイクロサービスの開発も行なっています。こちらはまだ道半ばで、プロジェクトとしてはちょうど半分ほど終わったところです。基本的には Rails で API server を書いているのですが、社内で広く使われるサービスである事から、「API の schema を定義し、その schema を反映した server の実装および client library の提供」を行うようにしています。
API schema をどう記述するかについては JSON Schema や OpenAPI などいくつか候補を考えたのですが、最終的には protocol buffer の「.proto ファイル」を利用することにしました。理由としては、1つは後々 grpc へ移行したいと考えているためそれが自然に出来るようにしたかったから、もう1つは .proto ファイルから OpenAPI format の JSON を出力できるため、OpenAPI のエコシステムを利用できたからです(e.g. Swagge UI)。「後々 grpc へ移行」と書いたように、今は grpc は利用しておらず、.proto で記述した format の JSON response を返す API server と、JSON response を parse して object へ mapping する client library を提供しています。
すぐに grpc を利用しないのは、現時点では HTTP/2 用の proxy を社内的に導入していないため、RPC に HTTP/2 を利用する基盤が整っていないからです。ただ、基盤の導入は進めようとしており、それが出来たタイミングで grpc へ移行したいと考えています。
「.proto ファイルで JSON response の schema を記述する」という取り組みは世の中的にほとんど行われていないので、自分で仕組みを整えながら開発を進めています。細かい部分は以下のように gem として切り出して OSS にしています。
- https://github.com/south37/field_mask_builder
- https://github.com/south37/field_mask_parser
- https://github.com/south37/pb_json_parser
- https://github.com/south37/fmcache
ただ、まだメインの仕組みについては公開できていません。メインの仕組みについても、後々整理して gem 化した上で、OSS にしたいと考えています。
仕事面の振り返りまとめ
インフラチームへ移った事で、働き方が大きく変わりました。 自分の興味の方向性ともあっており、良い変化だったと思います。
プライベート面の振り返り
プライベート面では、2017 年以降、OSS 活動に積極的になったと感じています。
きっかけは、2017 年 3 月に同僚や他社の知り合いと一緒に何気なく始めた Meguro.rb という地域 ruby コミュニティです。Meguro.rb は目黒近辺の会社で持ち回りでホストと会場提供をしながら LT と懇親会をしていて、最初は Wantedly、ドリコム、Viibar の3社だけで始めました。開催するうちに Livesense, ラクスル, アカツキ, freee, Quipper, CaSy, SUPERSTUDIO とホスト企業がどんどん増えていき、今ではホスト企業は 10 社になりました。
Meguro.rb を始めてから、LT 発表する機会が増え、発表があると発表のために何か成果を出そうというモチベーションが生まれました。それが、OSS 活動にもつながったと思います。
2017、2018 年に行った OSS 活動について
2017、2018 年に行った OSS 活動について、いくつかまとめておこうと思います。
以下はいくつかのプロジェクトに対して送った PR の一部です。
- https://github.com/ruby/ruby/pull/1626
- https://github.com/zilkey/active_hash/pull/156
- https://github.com/fluent/fluent-plugin-s3/pull/238
- https://github.com/motemen/ghq/pull/85
- https://github.com/goby-lang/goby-lang.github.io/pull/5
個人的には、一番上の「Ruby 本体に contribute 出来たこと」が一番嬉しかったです(PR 自体は Close されているが、それは Ruby の contribution の扱いの方針によるもので、変更自体は別の commit で取り込まれている)。やったのは、「Ruby の文字列補完(String Interpolation)のパフォーマンス改善」です。Ruby は自分が最も書き慣れており仕事でも使っている言語なので、そのベースの機能を改善出来たことはとても良い体験でした。また、String Interpolration はベーシックな機能なので、インパクトも大きいと感じてます。
この contribution は、Meguro.rb でも紹介しました。
その他、自作のツール/ライブラリも OSS としていくつか公開しました。
- https://github.com/south37/rucc
- https://github.com/south37/yard2steep
- https://github.com/south37/cf
- https://github.com/south37/parray
rucc と yard2steep は Meguro.rb で発表したのですが、反響があって嬉しかったです。
rucc は Ruby で書いた C 言語コンパイラです。作ったきっかけは、コンパイラの勉強の為に 8cc という rui314 さん作の C コンパイラのコードを読んでいた事です。読むだけでなく写経した方が勉強になると考え、さらにただ C から C へ写経しても面白くないので Ruby へ移植する事にしました。class 構造の整理や ruby の機能による書き換え、足りない機能の追加などはありますが、ロジックの大部分は 8cc を参考にしています。時間はかかりましたが、とても楽しかったです。
yard2steep は、Ruby コード中の YARD アノーテーションを元に静的型チェックを行うためのツールです。2018 年の RubyKaigi で紹介されたsteep という「Ruby で静的型チェックを行うためのツール」向けに、型定義ファイルを Ruby コード中の YARDアノーテーションを解析して生成します。steep の発表の際に、「YARD を利用したい」という声を聞いて試しに作ってみたものです。実は、RubyKaigi 中のコード懇親会というイベントでプロトタイプを作ったのが project を始めたきっかけで、手応えがあったのでそのまま作りきりました。最初は parser を手で書いていて、それが楽しかったです(正確な文法を反映していなかったので、後で Ripper ベースに書き換えてしまいましたが)。
OSS 活動についてまとめ
2017 年から、OSS 活動に以前に比べて積極的になり、OSS に対する心理的ハードルがどんどん下がっています。良い変化だったと思います。
英語学習について
2018年は英語に対するモチベーションが上がって、DMM 英会話をやってみたり、1つの本を Audible と Kindle 両方で買って英語音声を聞きながら英語文章を読んだり、映画を英語音声/英語字幕で観たりし始めました。
きっかけは色々ありましたが、仕事の関係で AWS re:Invent に行けたのも大きいです。自分にとって初の海外カンファレンスで、行列で隣り合った現地のアメリカ人と話したり、Pab Crawl でアメリカに長く住む中国人と話したり出来ました。実は行く数ヶ月前から細々と英語学習をしていたのですが、現地で会話してみて「丁寧にゆっくり話してもらわないと聞き取りが難しい」ことを痛感し、より英語学習へのモチベーションが上がりました。
これも、良い変化だったと思います。
まとめ
2017、2018 年の2年間を振り返ると、仕事面/プライベート面の両方でポジティブな変化がたくさんありました。 良い取り組みについては今後も継続して取り組んでいきたいですし、今後もさらにポジティブな変化を起こしていきたいです。
ISUCON7本戦に出場してきました
ISUCON7本戦に出場してきました。 結果はスコア0、最終的にはFAILで終わるという惨敗でした。
出場メンバーは会社の同僚との3人組で、予選と同様に他の2人がアプリケーション担当、僕がインフラ担当という役割分担をしました。
残念な結果となってしまい、とても悲しい気持ちなのですが、その思いを昇華させるためにも当日やったこと、考えていたことなどを記録しておこうと思います。 尚、悲しい結果に終わったこともあり、後悔や言い訳の多い見苦しい文章となっています。あらかじめご了承ください。
当日の流れ
10:00 にレギュレーションが閲覧できる様になりました。 しばらくは運営側でid, passwordの配布に時間がかかっており、コンテストが始まらなかったので、その間に3人でレギュレーションの読み合わせを行いました。単純に読んでいるだけでも、複雑なアプリケーションで単純な最適化は難しそうだなーという印象を受けました。
最初の30分は僕が各サーバーのセットアップ(git管理下に置く、各人がsshできる様にauthorized_keysを配置する、netdataの監視を仕込むなど)を行い、他2人がDBのデータの確認とアプリケーションのコードを読んでの概要把握を行いました。
その後、最適化の方法をざっくり話し合い、以下の方針を決めました。
- 方針1.
calc_status
が明らかにヘビーだから、なんらかの形でキャッシュする必要はありそう - 方針2. dbを触ってる部分はRedisに載せていくと良さそう
- 方針3. roomごとにスコア計算は完全に独立しているから、roomごとにhostを振り分けて、hostごとにデータを持つ様にしたらthroughputが出そう(この時点では、redisをhostごとに建ててそこにデータをstoreすることを考えていたので、redisの負荷が減りそうくらいのイメージで捉えていた)
- 方針4. static file配信はいつもの様にnginxでgzipで配信すると良さそう
今振り返るとこの時点での方針は悪くなかったと思うのですが、結果的には「方針4」以外は実装が思う様に進まず、ほぼmergeされる事はありませんでした。
自分が後悔している事の一つは、この時アプリケーションを読んでいた同僚から「ロックが必要になる」という事を言われた際に、ロックの方法についてもうちょっと自分もアプリケーションを読み込んだ上で議論すべきだったという事です。
自分はアプリケーションのコードは1行も読んでなくて、ただ「DBでロックを獲得するコードがある」と言われて議論した結果、DBを捨ててRedisを使うならRedisのincr
でロック獲得を記述できそうという結論になりました。この「Redisのincr
で実装したロック」は動作はしたのですが、毎回ロック獲得トライをRedisに問い合わせて行うというコストの高い動作の為か、スコアアップには繋がりませんでした。後で同僚からは「ロックはroom_nameごとに独立してかければ良かった」と聞かされて、そこまでの知識があればこの時点でroomごとに処理するprocessを分けて、ruby内部でmutexを活用しようという提案ができたかも。。。と後悔しています。
現実には、同僚の一人が「方針1」を、もう一人が「方針2」を進め、僕は「方針4」を行った後、便利スクリプト集(deployやログ閲覧、アクセスログ回収、各ホストの状態チェックなどをローカルのマシンからコマンド一発で実行できるもの)の整備やちょっとしたパラメータチューニング(pumaが10processで動いてmemoryを結構使っていたのでprocess数を減らしてthread数を増やしたりしてた)、2人とのペアプロなどを行なっていました。
「方針4」は一応スコアに効いて、9,600ぐらいになりました。しかし、それ以降僕らのチームのスコアが上がる事はありませんでした。
同僚が2人とも苦戦していたので、僕は定期的にペアプロでハマり解消を手伝っていました。
「方針1」を進めていた同僚が「途中まで出来た」といってベンチにかけてみたところ、スコアは上がらないどころか若干下がっている様に見えたので、その時はもう少し実装を進めてスコアを上げてからmergeしようという話になりました。結局、その後そのブランチがベンチを通る事はなく、彼は最後までそのブランチで作業し続けることになりました。(余談ですが、スコアが上がらなかったのは、おそらく前述の通りRedisをロックに利用していた為では無いかと疑っています)
「方針2」を進めていた同僚もRedisでロックをかけていて、こちらもベンチがなかなか通らずずっとハマっていました。最終的にロックを色々な箇所に設定する事でベンチは通ったみたいですが、それは終了の30分前であり、スコアも上がりませんでした。
(尚、「スコアが上がらなくてもmergeだけでもしておこう」といって「方針2」の一部はmergeされたのですが、最終スコアはFAILになってしまいました。不思議に思ってコードを見返すと、分かりやすく「引数が無いメソッドコール」というバグがありました。「ベンチ通ったからmergeした」と聞いたのに悲しい。。。)
結果としては、1日を通して有効な変更をほぼ入れられず、やりたいことが実現できないまま終了してしまったコンテストとなりました。
反省
自分の反省点はいろいろありますが、大きいのは以下の2つだと考えています。
1つ目は、自分がアプリケーションを全く読み込まなかったことです。「アプリケーションの仕様」を理解してないことから、途中で2人とペアプロする際も「コード的な間違いの指摘、修正」に終始してしまい、「アプリケーションの仕様の範囲で可能な限り効率的な実装を考える」という事が出来ませんでした。特に、今回の問題は効率的な実装をするにはスレッドプログラミングの知識が求められるものになっていたと思うので、自分の知識は本当はもっと活かせたんじゃ無いかと思っています。
2つ目は、自分がアプリケーションコードをほとんど書かなかった事です。これは1つ目の反省とも繋がるのですが、アプリケーションを読み込んでない状態では改善のアイディアを考えづらく、さらに残り時間が少なくなるほど「これからアプリケーションに着手する」といった手が取りづらくなり、結果としてほとんどアプリケーションを書きませんでした。アプリケーションがボトルネックというのは初めから分かっていたので、インフラ作業を捨てるというのも1つの選択肢だったように思います。
自分は普段はアプリケーションエンジニアとして仕事をしており、ISUCONは貴重な「インフラ知識をつけることができる、インフラオペレーションに慣れる事ができる」機会となっています。カーネルやミドルウェアのチューニングをしてマシンを効率的に利用したり、デプロイツールなどを自作して様々なタスクを自動化するのは僕にとって楽しくて、ISUCONではインフラ担当としての参加にこだわっていました。ただ、今回はそのこだわりが良く無い方向に働いたな、と感じています。
結論
今年は悔しい結果となったので、また来年頑張りたいです!!!
謝辞
運営の皆さん、お疲れ様でした。綺麗に整理されたコードで複雑なアプリケーションが実装されており、運営の皆さんの気合を感じました。取り組みがいのある、良い問題だったと思います。ありがとうございました。
ISUCON7予選をギリギリの18位で突破しました
会社の同僚と一緒にISUCON7予選に出場しました。 結果は18位、ギリギリでしたが何とか予選突破できました。事前にチームで練習したり、便利スクリプト集を仕込んだりしていたので、練習の成果がちゃんと発揮できてよかったです。
同僚が当日やったことについてブログを書いてくれました。余裕があれば、僕の観点からのISUCON7予選についてもブログ書きたいと思ってます。