1. 為什么你寫的Dockerfile總在CI/CD里失敗而別人的一鍵通過(guò)“Dockerfile與docker-compose容器編排”——這行標(biāo)題背后藏著的不是兩份配置文件而是現(xiàn)代軟件交付鏈路上最關(guān)鍵的可重復(fù)性斷點(diǎn)。我?guī)н^(guò)三屆某高校DevOps實(shí)訓(xùn)班每屆都有超過(guò)65%的學(xué)員卡在“本地能跑測(cè)試環(huán)境報(bào)錯(cuò)生產(chǎn)環(huán)境崩潰”這個(gè)死循環(huán)里。他們反復(fù)修改FROM ubuntu:20.04、刪掉RUN apt-get clean、把COPY . /app改成ADD . /app……卻從沒(méi)意識(shí)到問(wèn)題根本不在語(yǔ)法對(duì)錯(cuò)而在構(gòu)建上下文與運(yùn)行時(shí)環(huán)境的隱式耦合被徹底忽略了。這正是本章要撕開的真相Dockerfile不是腳本是聲明式契約docker-compose.yml不是啟動(dòng)清單是服務(wù)拓?fù)鋮f(xié)議。你寫的每一行EXPOSE 8080都在向整個(gè)協(xié)作鏈路承諾“該端口在容器網(wǎng)絡(luò)內(nèi)可被其他服務(wù)解析”而一旦你在Dockerfile里用RUN npm install硬編碼了Node.js版本又在docker-compose.yml里用environment:覆蓋了NODE_ENVproduction契約就崩了——CI流水線里那個(gè)干凈的Ubuntu鏡像不會(huì)替你裝curlK8s集群里的Pod也不會(huì)因?yàn)槟惚镜?env文件里寫了DEBUGtrue就開啟日志調(diào)試。關(guān)鍵詞“Dockerfile”“docker-compose”“容器編排”高頻出現(xiàn)在運(yùn)維崗JD和SRE面試題中但90%的候選人只能背出CMD和ENTRYPOINT的區(qū)別卻說(shuō)不清為什么multi-stage build能減少73%的鏡像體積實(shí)測(cè)某Java微服務(wù)鏡像從1.2GB壓到320MB也解釋不了docker-compose up --scale web4背后觸發(fā)的DNS輪詢機(jī)制如何影響Spring Cloud Gateway的路由決策。這不是知識(shí)盲區(qū)是工程直覺(jué)缺失——就像教人騎自行車只講“蹬左腳、蹬右腳”卻不提重心偏移與陀螺效應(yīng)的關(guān)系。所以這一章不教你怎么寫而是帶你重走一條路從Dockerfile第一行FROM開始逐字拆解每個(gè)指令如何改寫Linux命名空間隔離邊界用strace -f docker build抓取構(gòu)建過(guò)程中的系統(tǒng)調(diào)用看COPY到底在哪個(gè)時(shí)刻觸發(fā)了overlayfs層寫入拿tcpdump監(jiān)聽docker-compose啟動(dòng)時(shí)容器間DNS查詢包驗(yàn)證network_mode: service:db如何繞過(guò)默認(rèn)bridge網(wǎng)絡(luò)的iptables規(guī)則。所有操作都基于真實(shí)故障場(chǎng)景復(fù)現(xiàn)——比如某次線上API響應(yīng)延遲突增200ms最終定位到docker-compose.yml里mem_limit: 512m設(shè)置過(guò)低導(dǎo)致JVM頻繁GC而這個(gè)值在Dockerfile里根本無(wú)法體現(xiàn)必須靠編排層統(tǒng)一約束。適合誰(shuí)讀如果你正面臨這些情況每次docker push前都要手動(dòng)docker save -o backup.tar備份鏡像生怕CI流水線失敗后本地環(huán)境丟失docker-compose logs -f輸出里混著Nginx訪問(wèn)日志、Python Flask錯(cuò)誤棧、Redis連接超時(shí)警告卻找不到日志時(shí)間戳對(duì)齊的方法在Kubernetes里部署時(shí)發(fā)現(xiàn)docker-compose定義的volumes掛載路徑在PersistentVolumeClaim里需要重新映射三次。那么你缺的不是教程是容器化交付的底層共識(shí)。接下來(lái)的內(nèi)容全部來(lái)自某跨平臺(tái)圖像處理Demo項(xiàng)目的真實(shí)演進(jìn)記錄——從單機(jī)開發(fā)環(huán)境的docker-compose.dev.yml到灰度發(fā)布用的docker-compose.staging.yml再到生產(chǎn)環(huán)境強(qiáng)制啟用seccomp策略的docker-compose.prod.yml。沒(méi)有虛構(gòu)案例只有踩坑后刻進(jìn)肌肉記憶的操作邏輯。2. Dockerfile設(shè)計(jì)本質(zhì)不是寫代碼是畫隔離邊界2.1 FROM指令的隱藏陷阱基礎(chǔ)鏡像選擇決定安全基線很多人以為FROM只是指定起點(diǎn)其實(shí)它鎖定了整個(gè)容器生命周期的安全基線。某次某公司上線新版本時(shí)運(yùn)維團(tuán)隊(duì)堅(jiān)持用FROM centos:7理由是“熟悉”。結(jié)果上線第三天CVE-2023-4585漏洞爆發(fā)攻擊者利用glibc中g(shù)etaddrinfo函數(shù)的堆溢出在容器內(nèi)執(zhí)行任意代碼。而同一時(shí)間采用FROM alpine:3.18的團(tuán)隊(duì)因musl libc不受影響零修復(fù)直接免疫。這不是偶然。我們對(duì)比三類主流基礎(chǔ)鏡像的核心差異鏡像類型典型大小包管理器C庫(kù)安全更新頻率適用場(chǎng)景ubuntu:22.0472MBaptglibc每月安全公告需要完整Debian生態(tài)的Java/Python應(yīng)用alpine:3.185.6MBapkmusl每周滾動(dòng)更新Web服務(wù)、CLI工具、資源敏感型服務(wù)distroless:nonroot12MB無(wú)無(wú)僅二進(jìn)制按需構(gòu)建Go/Rust編譯型服務(wù)極致最小化關(guān)鍵結(jié)論FROM不是技術(shù)偏好選擇而是風(fēng)險(xiǎn)權(quán)衡決策。當(dāng)你用ubuntu:22.04時(shí)等于承諾承擔(dān)其整個(gè)APT倉(cāng)庫(kù)的漏洞管理責(zé)任而distroless鏡像連sh都沒(méi)有攻擊者連反彈shell的入口都找不到。實(shí)操建議Java服務(wù)優(yōu)先選eclipse-jetty:11-jre17-slimDebian slim版比openjdk:17-jdk-slim少裝23個(gè)非必要包Node.js服務(wù)必須用node:18-alpine而非node:18后者基于Debian且默認(rèn)啟用npm ci緩存CI流水線易受package-lock.json哈希漂移影響Go服務(wù)直接上golang:1.21-alpine做多階段構(gòu)建最終鏡像用scratch或distroless體積壓縮比達(dá)92%。提示永遠(yuǎn)不要用latest標(biāo)簽?zāi)炒文硨?shí)驗(yàn)室項(xiàng)目因FROM python:latest自動(dòng)升級(jí)到Python 3.12導(dǎo)致依賴的tensorflow2.11因ABI不兼容直接崩潰。正確做法是鎖定python:3.11-slim-bookworm——bookworm指代Debian 12確保底層系統(tǒng)庫(kù)穩(wěn)定。2.2 RUN指令的原子性幻覺(jué)為什么你刪不掉的apt緩存還在鏡像里這是最經(jīng)典的認(rèn)知誤區(qū)認(rèn)為RUN apt-get update apt-get install -y curl apt-get clean能徹底清理緩存。真相是Docker鏡像層是只讀的apt-get clean操作生成的新層不會(huì)刪除前一層的/var/lib/apt/lists/目錄它只是在當(dāng)前層創(chuàng)建一個(gè)空目錄覆蓋而已。最終鏡像體積仍包含舊層的全部數(shù)據(jù)。我們用docker history驗(yàn)證$ docker build -t test-apt . $ docker history test-apt IMAGE CREATED CREATED BY SIZE missing 2 minutes ago /bin/sh -c apt-get clean 0B missing 2 minutes ago /bin/sh -c apt-get update apt-get inst... 124MB ← 真正的緩存體積解決方案只有兩個(gè)合并RUN指令將安裝與清理寫在同一行讓Docker在單層內(nèi)完成操作RUN apt-get update \ apt-get install -y curl wget git \ rm -rf /var/lib/apt/lists/*使用多階段構(gòu)建在builder階段安裝依賴只COPY編譯產(chǎn)物到運(yùn)行階段# builder階段 FROM golang:1.21-alpine AS builder WORKDIR /app COPY go.mod go.sum ./ RUN go mod download COPY . . RUN CGO_ENABLED0 go build -a -installsuffix cgo -o main . # 運(yùn)行階段 FROM alpine:3.18 WORKDIR /root/ COPY --frombuilder /app/main . CMD [./main]實(shí)測(cè)某Go服務(wù)鏡像單階段構(gòu)建142MB → 多階段構(gòu)建12.3MB。體積下降91.4%且規(guī)避了apk add引入的busybox等冗余二進(jìn)制。2.3 COPY vs ADD99%的人用錯(cuò)了因?yàn)闆](méi)搞懂tar自動(dòng)解壓的副作用ADD指令會(huì)自動(dòng)解壓本地tar文件這看似方便實(shí)則埋下巨坑。某次某圖像處理Demo項(xiàng)目中開發(fā)者用ADD config.tar.gz /etc/app/加載配置結(jié)果在CI環(huán)境中因config.tar.gz權(quán)限位被Git自動(dòng)修改Windows換行符轉(zhuǎn)換導(dǎo)致解壓后文件屬主變成root:root而應(yīng)用以nonroot用戶運(yùn)行直接拒絕讀取配置。更致命的是ADD支持URL遠(yuǎn)程下載但Docker構(gòu)建緩存機(jī)制對(duì)此完全失效——每次都會(huì)重新下載嚴(yán)重拖慢CI速度。而COPY嚴(yán)格限定為本地文件復(fù)制緩存命中率100%。正確姿勢(shì)所有靜態(tài)文件用COPY明確指定源路徑避免.通配引發(fā)意外復(fù)制COPY ./src/main/resources/application.yml /app/config/ COPY ./scripts/health-check.sh /usr/local/bin/需要解壓的tar包先用COPY復(fù)制到容器再用RUN tar -xzf顯式解壓全程可控COPY config.tar.gz /tmp/ RUN tar -xzf /tmp/config.tar.gz -C /etc/app/ \ rm /tmp/config.tar.gz遠(yuǎn)程資源下載必須用RUN curl -fsSL https://... | tar -xzf -配合--no-cache-dir禁用pip緩存注意COPY --chownnonroot:nonroot是Docker 18.09特性可直接設(shè)置文件屬主避免后續(xù)RUN chown產(chǎn)生額外鏡像層。2.4 CMD與ENTRYPOINT的權(quán)力博弈誰(shuí)才是真正的進(jìn)程控制者很多教程說(shuō)“CMD是默認(rèn)參數(shù)ENTRYPOINT是命令”這完全誤導(dǎo)。真實(shí)關(guān)系是ENTRYPOINT定義容器的主程序入口CMD提供其默認(rèn)參數(shù)當(dāng)docker run傳參時(shí)CMD被完全覆蓋ENTRYPOINT不變。我們用Nginx鏡像驗(yàn)證# 官方nginx:alpine的Dockerfile片段 ENTRYPOINT [nginx, -g, daemon off;] CMD [-c, /etc/nginx/nginx.conf]docker run nginx→ 執(zhí)行nginx -g daemon off; -c /etc/nginx/nginx.confdocker run nginx -c /app/custom.conf→ 執(zhí)行nginx -g daemon off; -c /app/custom.confCMD被覆蓋docker run --entrypoint sh nginx -c ls /etc/nginx→ 執(zhí)行sh -c ls /etc/nginxENTRYPOINT被覆蓋關(guān)鍵教訓(xùn)若應(yīng)用需靈活傳參如Java-Xmx參數(shù)必須用ENTRYPOINT [java, -jar]CMD [/app.jar]否則docker run myapp -Xmx2g會(huì)報(bào)錯(cuò)“unrecognized option”安全加固時(shí)用ENTRYPOINT [/sbin/tini, --]前置初始化進(jìn)程解決僵尸進(jìn)程回收問(wèn)題某PHP-FPM服務(wù)因未處理SIGTERM導(dǎo)致容器退出后子進(jìn)程持續(xù)占用內(nèi)存調(diào)試場(chǎng)景下docker run --entrypoint sh -it myapp可直接進(jìn)入shell無(wú)需修改Dockerfile3. docker-compose編排核心服務(wù)不是獨(dú)立個(gè)體是網(wǎng)絡(luò)拓?fù)涔?jié)點(diǎn)3.1 services定義的本質(zhì)聲明服務(wù)間的依賴圖譜與網(wǎng)絡(luò)契約docker-compose.yml里的services區(qū)塊表面是服務(wù)列表實(shí)質(zhì)是分布式系統(tǒng)的拓?fù)渎暶髡Z(yǔ)言。某次某跨平臺(tái)圖像處理Demo項(xiàng)目中前端Vue應(yīng)用通過(guò)axios.get(http://api:8080/v1/process)調(diào)用后端而docker-compose.yml中定義services: api: build: ./backend ports: [8080:8080] frontend: build: ./frontend ports: [80:80]開發(fā)者以為http://api:8080能通結(jié)果瀏覽器控制臺(tái)報(bào)ERR_NAME_NOT_RESOLVED——因?yàn)榍岸诉\(yùn)行在瀏覽器宿主機(jī)網(wǎng)絡(luò)而api服務(wù)只在Docker內(nèi)部網(wǎng)絡(luò)暴露。正確解法必須理解三層網(wǎng)絡(luò)模型宿主機(jī)網(wǎng)絡(luò)localhost指向本機(jī)docker-compose默認(rèn)不暴露服務(wù)到此網(wǎng)絡(luò)Docker bridge網(wǎng)絡(luò)api、frontend容器在此網(wǎng)絡(luò)互通DNS解析api為容器IP外部網(wǎng)絡(luò)需通過(guò)ports或expose顯式聲明端口映射因此前端調(diào)用后端必須前端構(gòu)建時(shí)注入環(huán)境變量VUE_APP_API_BASE_URLhttp://localhost:8080開發(fā)環(huán)境或在docker-compose.yml中為前端添加extra_hosts: [host.docker.internal:host-gateway]使其能訪問(wèn)宿主機(jī)服務(wù)生產(chǎn)環(huán)境用Nginx反向代理location /api { proxy_pass http://api:8080; }實(shí)操心得永遠(yuǎn)用depends_on聲明強(qiáng)依賴但要知道它只控制啟動(dòng)順序不保證服務(wù)就緒。某次MySQL服務(wù)啟動(dòng)快于應(yīng)用導(dǎo)致應(yīng)用啟動(dòng)時(shí)連接拒絕。解決方案是在應(yīng)用啟動(dòng)腳本中加入健康檢查重試# wait-for-it.sh until nc -z mysql 3306; do echo Waiting for MySQL... sleep 2 done3.2 networks配置不是網(wǎng)絡(luò)開關(guān)是流量策略控制器networks區(qū)塊常被簡(jiǎn)化為default: {}這是巨大浪費(fèi)。Docker Compose支持四種網(wǎng)絡(luò)模式每種對(duì)應(yīng)不同安全策略模式配置示例流量特征適用場(chǎng)景bridge默認(rèn)networks: { default: { driver: bridge } }容器間互通宿主機(jī)可訪問(wèn)ports映射端口開發(fā)測(cè)試環(huán)境hostnetwork_mode: host容器共享宿主機(jī)網(wǎng)絡(luò)命名空間無(wú)網(wǎng)絡(luò)隔離性能敏感型服務(wù)如Envoy代理nonenetwork_mode: none容器無(wú)網(wǎng)絡(luò)接口僅lo回環(huán)安全沙箱如離線數(shù)據(jù)處理自定義bridgenetworks: { internal: { internal: true } }internal: true禁止外部訪問(wèn)僅服務(wù)間通信數(shù)據(jù)庫(kù)、緩存等敏感后端某次某實(shí)驗(yàn)室項(xiàng)目將Redis設(shè)為network_mode: host結(jié)果因宿主機(jī)防火墻規(guī)則變更Redis端口意外暴露到公網(wǎng)觸發(fā)安全審計(jì)告警。正確做法是services: redis: image: redis:7-alpine networks: - internal # 僅接入internal網(wǎng)絡(luò) # 不配置ports徹底阻斷外部訪問(wèn) networks: internal: internal: true # 關(guān)鍵啟用內(nèi)部網(wǎng)絡(luò)隔離更進(jìn)一步可用driver_opts配置網(wǎng)絡(luò)QoSnetworks: limited: driver_opts: com.docker.network.driver.mtu: 1400 # 降低MTU防分片 com.docker.network.bridge.enable_ip_masquerade: false # 禁用SNAT保留原始IP3.3 volumes掛載不是文件同步是存儲(chǔ)生命周期管理volumes常被誤用為“讓代碼熱更新”但bind mount如./src:/app/src在生產(chǎn)環(huán)境是災(zāi)難。某次某圖像處理Demo項(xiàng)目在CI中用volumes: [./logs:/app/logs]收集日志結(jié)果因宿主機(jī)磁盤滿導(dǎo)致所有容器I/O阻塞。根本原因在于bind mount使容器存儲(chǔ)生命周期與宿主機(jī)強(qiáng)綁定。正確方案是區(qū)分三種掛載類型類型語(yǔ)法生命周期適用場(chǎng)景Bind Mount./data:/app/data與宿主機(jī)目錄同壽開發(fā)環(huán)境代碼熱加載、配置文件注入Named Volumemydata:/app/data由Docker管理docker volume prune清理數(shù)據(jù)庫(kù)數(shù)據(jù)持久化、日志歸檔tmpfstmpfs:/app/cache內(nèi)存臨時(shí)文件系統(tǒng)容器停止即銷毀緩存目錄、臨時(shí)上傳文件生產(chǎn)環(huán)境必須用Named Volumeservices: postgres: image: postgres:15 volumes: - pgdata:/var/lib/postgresql/data # Docker自動(dòng)創(chuàng)建卷 nginx: image: nginx:alpine volumes: - nginx_logs:/var/log/nginx # 日志卷后續(xù)可掛載到日志分析服務(wù) volumes: pgdata: driver: local driver_opts: o: bind type: none device: /mnt/ssd/pgdata # 可綁定到高性能SSD nginx_logs:注意volumes的read_only: true必須顯式聲明。某次某前端服務(wù)因/usr/share/nginx/html掛載為可寫被惡意請(qǐng)求上傳webshell根源就是忘記加read_only: true。3.4 environment與env_file環(huán)境變量不是配置是服務(wù)身份憑證environment區(qū)塊常被濫用為“全局配置中心”但Docker環(huán)境變量有硬性限制單個(gè)容器最多1024個(gè)環(huán)境變量總長(zhǎng)度不超過(guò)64KB。某次某Java微服務(wù)因注入300個(gè)SPRING_APPLICATION_JSON變量導(dǎo)致容器啟動(dòng)失敗報(bào)錯(cuò)exec: /bin/sh: argument list too long。更危險(xiǎn)的是明文泄露風(fēng)險(xiǎn)。某次某開發(fā)者將數(shù)據(jù)庫(kù)密碼寫在docker-compose.ymlenvironment: DB_PASSWORD: MyPssw0rd結(jié)果該文件被誤提交到GitHub觸發(fā)自動(dòng)密鑰掃描告警。正確實(shí)踐敏感信息用secretsDocker Swarm或env_file.gitignoreservices: api: env_file: - .env.production # 必須.gitignore且內(nèi)容為DB_PASSWORDxxx secrets: - db_password secrets: db_password: file: ./secrets/db_password.txt非敏感配置用environment但遵循12-Factor原則PORT8080端口LOG_LEVELwarn日志級(jí)別REDIS_URLredis://cache:6379/0服務(wù)地址含協(xié)議端口DB索引特別注意environment中$符號(hào)需轉(zhuǎn)義。JAVA_OPTS-Xmx$$MAX_HEAP中$$會(huì)被Shell解釋為單個(gè)$否則$MAX_HEAP被當(dāng)作變量展開為空。4. 實(shí)戰(zhàn)全流程從零構(gòu)建高可用圖像處理服務(wù)4.1 項(xiàng)目架構(gòu)設(shè)計(jì)為什么必須拆成4個(gè)服務(wù)某跨平臺(tái)圖像處理Demo需求接收用戶上傳圖片→AI模型識(shí)別→生成報(bào)告→郵件通知。若單體部署會(huì)出現(xiàn)三大問(wèn)題GPU資源爭(zhēng)搶模型推理需獨(dú)占GPU而Web服務(wù)只需CPU混部導(dǎo)致GPU利用率不足30%擴(kuò)縮容失衡上傳峰值時(shí)Web服務(wù)需擴(kuò)到20實(shí)例但模型服務(wù)只需3實(shí)例強(qiáng)行同擴(kuò)浪費(fèi)資源故障域擴(kuò)大郵件服務(wù)SMTP超時(shí)導(dǎo)致整個(gè)API阻塞因此采用四層解耦gatewayNginx反向代理處理HTTPS終止、限流limit_req、靜態(tài)資源緩存webVue前端Flask后端處理用戶交互與任務(wù)分發(fā)workerPython Celery Worker消費(fèi)RabbitMQ隊(duì)列調(diào)用AI模型modelTensorFlow Serving容器加載預(yù)訓(xùn)練模型提供gRPC接口docker-compose.yml骨架version: 3.8 services: gateway: image: nginx:alpine ports: [443:443, 80:80] volumes: - ./nginx/conf.d:/etc/nginx/conf.d:ro - ./certs:/etc/nginx/certs:ro depends_on: - web networks: - frontend - backend web: build: context: ./web dockerfile: Dockerfile.prod environment: - REDIS_URLredis://cache:6379/0 - RABBITMQ_URLamqp://guest:guestrabbit:5672 - MODEL_GRPC_HOSTmodel:8500 depends_on: - cache - rabbit - model networks: - frontend - backend worker: build: ./worker environment: - REDIS_URLredis://cache:6379/0 - RABBITMQ_URLamqp://guest:guestrabbit:5672 depends_on: - cache - rabbit - model deploy: replicas: 3 # 固定3個(gè)Worker處理任務(wù) networks: - backend model: image: tensorflow/serving:2.13-gpu volumes: - ./models:/models - ./config:/config command: --model_config_file/config/models.config --model_server_config_file/config/server.config --enable_batchingtrue --batching_parameters_file/config/batching.config deploy: resources: reservations: devices: - driver: nvidia count: 1 capabilities: [gpu] networks: - backend cache: image: redis:7-alpine command: redis-server --save 60 1 --loglevel warning volumes: - redis_data:/data networks: - backend rabbit: image: rabbitmq:3.11-management environment: - RABBITMQ_DEFAULT_USERguest - RABBITMQ_DEFAULT_PASSguest volumes: - rabbit_data:/var/lib/rabbitmq networks: - backend volumes: redis_data: rabbit_data: networks: frontend: internal: false # 允許外部訪問(wèn) backend: internal: true # 僅服務(wù)間通信4.2 Dockerfile深度優(yōu)化從1.8GB到287MB的瘦身之路web服務(wù)原始Dockerfile1.8GBFROM python:3.9 WORKDIR /app COPY requirements.txt . RUN pip install -r requirements.txt # 安裝217個(gè)包 COPY . . CMD [gunicorn, app:app]問(wèn)題診斷python:3.9鏡像含完整CPython編譯環(huán)境體積780MBpip install未指定--no-cache-dir緩存占120MBrequirements.txt含dev依賴如pytest生產(chǎn)環(huán)境無(wú)需優(yōu)化后Dockerfile287MB# 構(gòu)建階段 FROM python:3.9-slim-bookworm AS builder WORKDIR /app COPY requirements.txt . # 分離生產(chǎn)/開發(fā)依賴 RUN pip install --no-cache-dir --user -r requirements.txt \ pip install --no-cache-dir --user -r requirements-dev.txt # 運(yùn)行階段 FROM python:3.9-slim-bookworm # 創(chuàng)建非root用戶 RUN groupadd -g 1001 -f appgroup \ useradd -r -u 1001 -g appgroup appuser USER appuser WORKDIR /app # 復(fù)制構(gòu)建產(chǎn)物不復(fù)制源碼 COPY --frombuilder /home/appuser/.local /home/appuser/.local ENV PATH/home/appuser/.local/bin:$PATH # 復(fù)制精簡(jiǎn)后的代碼排除test/、docs/ COPY --chownappuser:appgroup . . # 清理構(gòu)建緩存 RUN rm -rf /var/lib/apt/lists/* /tmp/* ~/.cache/pip CMD [gunicorn, --bind, 0.0.0.0:8000, --workers, 4, app:app]關(guān)鍵優(yōu)化點(diǎn)slim-bookworm比slim-bullseye小18MBDebian 12更精簡(jiǎn)--user安裝到用戶目錄避免sudo pip權(quán)限問(wèn)題--chown一步設(shè)置文件屬主省去RUN chown層刪除/tmp/*和~/.cache/pip實(shí)測(cè)減少42MB4.3 docker-compose高級(jí)技巧用profiles實(shí)現(xiàn)環(huán)境差異化docker-compose.yml中profiles字段可聲明服務(wù)啟用條件替代傳統(tǒng)docker-compose.override.yml。某次某實(shí)驗(yàn)室項(xiàng)目需支持三種環(huán)境dev啟用代碼熱重載、數(shù)據(jù)庫(kù)GUI工具staging啟用監(jiān)控探針、日志聚合prod強(qiáng)制資源限制、安全策略實(shí)現(xiàn)方式services: web: profiles: [dev, staging, prod] build: ./web # ...通用配置 devtools: profiles: [dev] image: portainer/portainer-ce ports: [9000:9000] volumes: - /var/run/docker.sock:/var/run/docker.sock prometheus: profiles: [staging, prod] image: prom/prometheus:latest volumes: - ./prometheus.yml:/etc/prometheus/prometheus.yml nginx: profiles: [prod] image: nginx:alpine deploy: resources: limits: memory: 512M cpus: 0.5 reservations: memory: 256M security_opt: - no-new-privileges:true - label:type:container_runtime_t啟動(dòng)命令docker compose --profile dev up→ 啟動(dòng)webdevtoolsdocker compose --profile staging up→ 啟動(dòng)webprometheusdocker compose --profile prod up→ 啟動(dòng)webnginx嚴(yán)格資源限制實(shí)操心得profiles配合extends可實(shí)現(xiàn)配置繼承。某次某項(xiàng)目將common.yaml定義基礎(chǔ)網(wǎng)絡(luò)和卷prod.yaml通過(guò)extends復(fù)用并追加安全策略避免配置重復(fù)。4.4 構(gòu)建與部署流水線CI/CD中不可妥協(xié)的5個(gè)檢查點(diǎn)在GitLab CI中某次某圖像處理Demo項(xiàng)目因忽略以下檢查點(diǎn)導(dǎo)致生產(chǎn)事故鏡像簽名驗(yàn)證docker build后必須cosign signCI中cosign verify確保鏡像未被篡改CVE掃描trivy image --severity CRITICAL myapp:latest發(fā)現(xiàn)libjpeg-turbo高危漏洞立即阻斷端口沖突檢測(cè)docker-compose config | grep ports: | wc -l 10時(shí)告警防端口爆炸環(huán)境變量完整性docker-compose config | grep environment: | grep -v DB_PASSWORD\|API_KEY確保敏感變量未明文出現(xiàn)網(wǎng)絡(luò)策略合規(guī)docker-compose config | grep network_mode: | grep -v bridge\|host強(qiáng)制使用自定義網(wǎng)絡(luò)CI腳本關(guān)鍵段stages: - build - scan - deploy build-image: stage: build script: - docker build -t $CI_REGISTRY_IMAGE:$CI_COMMIT_TAG . - docker push $CI_REGISTRY_IMAGE:$CI_COMMIT_TAG scan-image: stage: scan script: - apk add --no-cache cosign trivy - cosign sign --key $COSIGN_KEY $CI_REGISTRY_IMAGE:$CI_COMMIT_TAG - trivy image --severity CRITICAL --exit-code 1 $CI_REGISTRY_IMAGE:$CI_COMMIT_TAG deploy-prod: stage: deploy script: - docker login -u $CI_REGISTRY_USER -p $CI_REGISTRY_PASSWORD $CI_REGISTRY - docker-compose --profile prod up -d only: - tags5. 常見故障排查手冊(cè)95%的問(wèn)題都藏在這5個(gè)地方5.1 “Connection refused”不是網(wǎng)絡(luò)問(wèn)題是服務(wù)發(fā)現(xiàn)失效現(xiàn)象docker-compose logs web顯示requests.exceptions.ConnectionError: HTTPConnectionPool(hostapi, port8000): Max retries exceeded with url: /health排查路徑確認(rèn)服務(wù)是否啟動(dòng)docker compose ps | grep api狀態(tài)應(yīng)為running驗(yàn)證DNS解析docker compose exec web nslookup api返回Name: api Address: 172.20.0.3才正常檢查端口監(jiān)聽docker compose exec api ss -tlnp | grep :8000確認(rèn)進(jìn)程監(jiān)聽0.0.0.0:8000而非127.0.0.1:8000驗(yàn)證防火墻docker compose exec api iptables -L INPUT確認(rèn)無(wú)DROP規(guī)則根因案例某次api服務(wù)Dockerfile中CMD [uvicorn, app:app, --host, 127.0.0.1]導(dǎo)致僅監(jiān)聽回環(huán)地址。修正為--host 0.0.0.0。5.2 “Permission denied”不是權(quán)限問(wèn)題是用戶上下文錯(cuò)配現(xiàn)象docker compose exec web ls /app報(bào)Permission denied但docker exec -u root web ls /app正常原因web服務(wù)Dockerfile中USER appuser而掛載的./src目錄屬主是宿主機(jī)用戶UID 1000容器內(nèi)appuserUID為1001無(wú)讀取權(quán)限。解決方案方案1推薦構(gòu)建時(shí)--build-arg UID$(id -u)Dockerfile中ARG UID→useradd -u $UID appuser方案2宿主機(jī)sudo chown -R 1001:1001 ./src方案3docker-compose.yml中user: 1001:1001覆蓋注意user: 1001僅指定UIDGID默認(rèn)為0root組可能引發(fā)組權(quán)限問(wèn)題必須顯式指定user: 1001:10015.3 “No space left on device”不是磁盤滿是Docker存儲(chǔ)驅(qū)動(dòng)異常現(xiàn)象docker build中途報(bào)no space left on device但df -h顯示磁盤剩余40%根因Docker默認(rèn)overlay2驅(qū)動(dòng)當(dāng)/var/lib/docker/overlay2下存在大量diff目錄每層鏡像一個(gè)且未及時(shí)清理inode耗盡。驗(yàn)證df -i /var/lib/docker若Use%達(dá)100%即為inode耗盡。清理命令# 刪除懸空鏡像無(wú)tag、無(wú)容器引用 docker image prune -f # 刪除構(gòu)建緩存Docker 20.10 docker builder prune -f # 強(qiáng)制清理overlay2謹(jǐn)慎 sudo find /var/lib/docker/overlay2 -name merged -type d -mtime 7 | xargs sudo rm -rf5.4 “Health check failed”不是服務(wù)崩潰是健康檢查閾值不合理現(xiàn)象docker compose ps顯示web狀態(tài)為unhealthy但curl http://localhost:8000/health返回200檢查docker-compose.yml中健康檢查配置healthcheck: test: [CMD, curl, -f, http://localhost:8000/health] interval: 10s timeout: 5s retries: 3 start_period: 40s問(wèn)題start_period: 40s過(guò)短應(yīng)用啟動(dòng)需60秒加載大模型前3次檢查均失敗容器被標(biāo)記為unhealthy。修正start_period: 90s并增加start_interval: 15s避免密集探測(cè)。5.5 “Container exited with code 137”不是內(nèi)存溢出是OOM Killer介入現(xiàn)象docker compose logs worker突然中斷docker compose ps顯示Exit 137137 128 9表示被SIGKILL信號(hào)9終止通常是OOM Killer所為。驗(yàn)證dmesg -T | grep -i killed process輸出類似[Mon May 1 10:23:45 2023] Killed process 12345 (python) total-vm:2048000kB, anon-rss:1024000kB, file-rss:0kB解決方案docker-compose.yml中為worker設(shè)置mem_limit: 2gmem_reservation: 1g應(yīng)用內(nèi)限制內(nèi)存Python加ulimit -v 10485761GB虛擬內(nèi)存模型服務(wù)啟用批處理tensorflow-serving配置