與收入息息相關，美團外賣廣告系統架構解密

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"前幾天無意刷到了 2021 Q2 互聯網公司廣告的營收榜單，驚訝於廣告業務的增長速度和進化能力，也對這背後的技術起了興趣，支撐如此龐大業務量和數據背後用了什麼技術？如何建立起穩定的廣告系統架構支撐？中間走過哪些坑，又解決什麼難題等等？帶着一系列疑問，我們採訪到了 ArchSummit 全球架構師峯會的講師--美團資深技術專家林樂彬，來解答這些問題。此外，他還會在 "},{"type":"link","attrs":{"href":"http:\/\/ArchSummit 深圳 2021","title":null,"type":null},"content":[{"type":"text","text":"ArchSummit 深圳 2021"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" 上帶來"},{"type":"link","attrs":{"href":"https:\/\/archsummit.infoq.cn\/2021\/shenzhen\/presentation\/4138","title":null,"type":null},"content":[{"type":"text","text":"《美團外賣廣告系統平臺化架構實戰》"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"的演講分享，感興趣的同學不要錯過。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"美團廣告系統架構探索與實踐"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"InfoQ：您可以先給大家簡單介紹一下廣告系統平臺架構的設計思路和構建過程嗎?"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬：架構是爲業務服務的"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，設計思路就是如何通過合理的架構設計保證業務的快速發展。即相對業務來說，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"架構保持適度的前瞻性但不能過渡設計"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"。因此我們架構構建的過程，會根據業務發展的階段特點所制定。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"我們的廣告系統是在 2016 年開始建設的，前期業務試水階段，我們通過 "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"MVP 方案"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"幫助業務快速拿到收益；之後，隨着業務體量的變大，主要體現在 C 端流量和 B 端的供給大幅增加，主要的工作是提升架構可擴展性，解決可能出現的系統瓶頸。在後來，隨着新老業務的不斷髮展，產品需求越來越多，研發人力是有限的，這個時候，我們開展了以提升研發效率爲目標的平臺化階段工作。"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"貫穿始終的是穩定性、性能優化、資源利用率提升的不斷建設"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"InfoQ：我們在構建如此龐大的廣告架構系統時，主要走過哪些坑，又解決了哪些技術痛點？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬："},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"遇到的問題，解決的技術痛點有很多，主要有以下幾點："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"1. 前期的數據同步系統比較簡單，廣告的更新數據，增量更新採用 B 端直接發送數據到下游服務，而全量更新則複用增量更新的方式更新數據。後面隨着業務體量的變大，在擴展性和穩定性都出現了一些問題，比如更新物料多了，積壓現象明顯等。此外每次服務發版的時候，切換流程都比較繁瑣，很容易出問題。後面我們借鑑 "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"ESB "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"的總線思路，自研了一套"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"流批一體"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"的數據通路平臺。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"2. 爲了提升迭代效率，我們開展了平臺化項目，在實現中，我們會通過 "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"DAG "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"的調度框架替代原先程序自身調度邏輯。新的框架帶來了便利性同時，由於引入了線程池的調度，性能會有一定的折損。我們通過"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"無鎖化調度"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"以及"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"線程模型調優"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"等方案基本保證了與原生的性能持平。此外在某些服務，由於業務邏輯十分複雜，人工很難給出最好的調度關係，因此無法達到最佳的性能，我們基於數據驅動的思想，實現了邏輯調用的自動化方案，從而極大的提升了業務的性能。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"3. 在算法架構方向，算法的迭代會經過“訓練樣本構建，模型訓練，評估，特徵上線，模型推送以及線上推理” 一系列過程，由於各個環節割裂，人工參與度高，以及中間出現大量重複開發，會導致算法迭代實驗週期很長，我們通過一站式的算法平臺解決了算法迭代的痛點。此外在線推理中，隨着模型複雜度增加以及體量變大，性能遇到較大挑戰，導致無法正常迭代，我們通過搭建 Workflow 的預估框架，異構計算等方案提升性能，保證算法持續迭代。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"InfoQ：目前我們廣告架構系統比較大的挑戰有哪些?"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬："},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"說到目前比較大的挑戰，首先，我們支持的業務線比較多，包括 feed 流形式的列表廣告，針對泛 ka 的展示廣告，搜索廣告 以及遊戲場景的創新廣告 等主要產品線，以及對應的 20+ 個細分的業務場景，所有這些業務的引擎架構都收口在我們團隊。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"這些業務場景相互之間邏輯差異很大，由於發展階段不同，因此業務的訴求也會有很大的差異。如何通過技術能力幫助各自業務快速發展，以及如何通過技術讓不同業務線之間相互賦能都是比較大的挑戰。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"其次，算法的日益複雜，以及產品的需求多樣化，導致耗時越來越多。而在上游超時一定的情況下，如何通過合理的架構設計與極致的性能優化，從而更好的爲業務提供服務也是重中之重。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"InfoQ：架構構建的過程中，在一些技術方向上我們做過哪些取捨，原因是什麼，您可以從技術的角度簡單講解一下嗎?"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬：快速支撐業務的思路還是技術平臺思路"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"。這是我們整體做過一個比較大的取捨，對於這個問題，我的思考是要根據業務的發展階段來決定，這也是我們架構構建的過程。在業務的早期，規模還比較小，業務希望快速拿到收益，這個時候我們會把有限的人力投入到快速支撐業務上。而當業務發展到了一定的規模，以及業務線逐漸增多的時候，我們會啓動技術平臺的思路建設，一方面通過技術能力的深入來解決規模化帶來的問題，同時也通過平臺化的能力提升多個業務線的功能複用度，幫助更多業務線提升迭代效率。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"此外在一些技術細節上我們也做過一些取捨，這裏舉一個外賣特色的例子：在廣告檢索方面，外賣廣告有着典型的 LBS"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":" "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"的配送範圍限制，因此針對商家的召回場景，因爲數量不會特別多，在架構設計的時候，根據業務的特點，在召回階段我們沒有采用傳統的正倒排的召回模式，而是採用了上游服務進行商家 LBS 範圍篩選，在召回階段只做非廣告商家過濾的方式。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"InfoQ：在高可靠性和可擴展性的架構中，我們如何處理系統可靠性和可擴展性等問題?"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬："},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"可靠性主要在以下兩個層面做了一些工作。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"1. 運營機制層面，我們針對日常，事前（事故發生前），事中（事故發生時候），事後 （事故發生後）等幾個階段分別建立了一套成體系的穩定性建設方案。在“日常”主要是一些基本能力的建設，包括流程規範、性能優化、日常演練等等；在事前做好各種力度監控，保證有錯誤可以第一時間發現；在事中，做好各種應急方案，如降級，回滾等；在事後，做好事後總結，做到可以舉一反三。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"2. 在具體的落地動作中，映射到架構層面，也做了很多。我這裏簡單講一下我們接下來要做的一個事情，"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"智能降級的方案。"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"將我們之前在智能算力沉澱的能力（可以參考我們之前在美團公衆號的文章-"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#333333","name":"user"}}],"text":"美團外賣廣告智能算力的探索與實踐"},{"type":"text","text":"），用到穩定性上面，即如果出現流量突增的時候，可以做到根據流量的價值，不同的流量選擇不同的降級方案，從而使在降級過程中，收入損失最小。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"可擴展性對於廣告系統來說，常見的就是兩種場景，一種是廣告請求，一種是廣告更新。前者是讀場景，後者是寫場景。針對廣告請求場景我們的可擴展性設計主要在兩個層級上，一個是針對服務，一個是針對邏輯功能。前者主要是通過服務的拆分的方式，即：針對流量，功能，存儲進行了拆分。後者邏輯功能部分主要通過平臺化中框架和組件的思想來實現解耦以及提升擴展性。針對廣告更新，我們主要是通過消息總線的方式來進行雙方的解耦，來達到可擴展的效果。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"不同業務場景帶來的架構調整挑戰"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"InfoQ：隨着業務的發展，每天廣告的請求量很大，流量分佈時間段也不平均，同時要求低延時響應，我們如何來處理這種高峯值爆發的情況?"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬："},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"外賣的業務，有着典型的潮汐效應，每天的流量高峯主要集中在午高峯和晚高峯期間。其實線上的機器數量也是按照高峯期的峯值進行預估的，正常情況是沒有問題。但是如果出現活動流量突增的時候，我們的"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"處理"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#212121","name":"user"}},{"type":"strong"}],"text":"原則是減少損失與影響範圍，並儘快進行恢復。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#212121","name":"user"}}],"text":"具體來說，減少損失指減少單次流量的損失，我們主要的思想是針對這次流量進行降級，我們會從服務，功能，數據，流程 多個維度進行降級。減少影響範圍指減少不同流量的相互影響，主要的思想是限流和隔離，其中針對限流，主要依賴公司的限流組件能力，可以實現機器，集羣，自定義功能等多維度的限流能力；隔離，主要是通過業務線間的部署隔離進行解決，防止不同業務線間的流量相互影響。此外也會利用公司容器動態擴容的能力，快速擴容，進行快速恢復。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"InfoQ：如何做到儘量不影響用戶體驗的情況下，提高廣告點擊率和收入，減少用戶流失?我們通過什麼樣的技術來處理商家、用戶和平臺之間的利益平衡，維持廣告業務的健康可持續發展？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬："},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"用戶、商家、平臺三方優化的問題，是廣告中的核心問題，我們通常有兩個階段，一是從三個方面分別去做優化，做大蛋糕，例如去優化用戶體驗、去優化商家的 "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"ROI "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，二是通過多目標優化的設計，去合理的平衡三方的收益，例如通過多目標 Rank Socre 優化的方式進行三方收益的合理優化。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"InfoQ：提到深度學習，目前深度學習技術在美團外賣廣告系統架構中還有哪些應用場景，具體實踐情況如何?"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬："},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"深度學習在美團外賣廣告的應用主要可以分成兩個方面，一是偏推薦的部分，從召回、粗排、搜索廣告裏的相關性優化，精排中的 CTR、CVR、客單價預估模型的迭代；二是偏廣告機制的部分，例如智能出價、智能創意、深度拍賣模型等等。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"國內外廣告系統架構的發展"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"InfoQ：前些年 DSP 在國內興起時很火爆，很多人對 DSP 期待也很高，現在來看並沒有達到人們的預期高度，您覺得其中主要的問題是什麼?"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬："},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"我覺得主要有以下幾點原因："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"1. 出於隱私以及利益等原因，數據不夠開放，導致媒體、客戶、平臺之間的數據割裂，限制各方策略優化空間；"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"2. 媒體更希望將優質的流量優先給自有客戶，因此對於 DSP 來說，和媒體的自有廣告主競爭的時候，不是絕對的公平；"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"3. DSP 的出現，降低了商家投放門檻，由於 App 的使用偏集中化，因此流量主要集中在頭部的一些媒體，因此會導致優質流量競爭十分激烈，從而擡升流量成本。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"InfoQ：最後聊點別的，我們知道微軟、Google 等國外公司在廣告領域做了很多創新，有哪些是值得我們國內廠商去學習的嗎，我們現在的優勢或不足又有哪些?"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬："},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在廣告領域，國外幾家頭部公司做出了很多具有影響力的工作。例如微軟在召回粗排中的 "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"DSSM"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"（Deep Structured Semantic Model）和 "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"YouTube"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"DNN "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"的"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"經典雙塔"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" + "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"ANN "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"範式，Google 精排 "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"Wide & Deep 模型"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"推動了深度學習在推薦廣告中的大規模落地，同時 "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"GSP "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"和 "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"VCG "},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"也已成爲了廣告競價機制的標準方式。拋開具體技術細節，國外幾家公司從業務實踐出發，敢於結合先進技術進行 0 到 1 的開創探險精神是最值得我們去關注的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"相比較而言，國內互聯網起步較晚，很長一段時間都是一個追隨者的角色，但近些年國內互聯網業務迅猛發展，產品技術不斷創新，在很多技術上已經大有趕超之勢。總結來看，國內產品模式的創新，海量的互聯網用戶，科技人才的湧現已成爲國內互聯網產品技術進一步發展的最大優勢。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"採訪嘉賓介紹"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"林樂彬"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，資深美團技術專家，2016 年加入美團，目前是美團外賣廣告引擎架構負責人，曾前後任職騰訊、百度，有 10 年以上互聯網廣告相關經驗。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"在 11 月 12-13 日舉辦的 "},{"type":"link","attrs":{"href":"https:\/\/archsummit.infoq.cn\/2021\/shenzhen\/","title":null,"type":null},"content":[{"type":"text","text":"ArchSummit 深圳 2021 全球架構師峯會"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"中，林樂彬老師將在【"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"高可用高性能業務架構】"},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"中分享“美團外賣廣告引擎平臺化實踐”。除了對廣告架構的介紹外，還會通過現狀分析、業務場景、落地案例深度解讀廣告架構實踐。更多精彩內容請掃碼查看。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/be\/bf\/be10d0b1a934bed3e95826dd7a4fa3bf.png","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}},{"type":"strong"}],"text":"關聯閱讀："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/article\/9jYdCOGBkoioO6zm1fVG","title":null,"type":null},"content":[{"type":"text","text":"Twitter 廣告平臺實時計費系統的架構增強之道"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/article\/ANNelGBb2eySlThMYqiv","title":null,"type":null},"content":[{"type":"text","text":"大規模圖算法在京東廣告的實踐"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/article\/jw59kl5ap5308ucvp1wm","title":null,"type":null},"content":[{"type":"text","text":"如何搭建批流一體大數據分析架構？"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/article\/939qLvUZxtLGModQUs2v","title":null,"type":null},"content":[{"type":"text","text":"面向下一代的粗排排序系統 COLD"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]}]}