Databricks 12 Services 部署教程
4 大类 × 18 个服务 × 时间 + 难度

0. 30 秒结论

你的问题	答案
Databricks 有多少服务可用?	12+ 个核心服务 跨 4 大类 (Data Access / ML Lifecycle / AI Agent / 共用)
最快上手的服务?	Databricks SQL (5-10 min) — Serverless warehouse create + Photon 加速
部署时间最长?	Feature Store (60 min) + Custom Agents (45-60 min) — 都需要懂 data model + MLflow tracing
跨境支付 use case?	RAG + Custom Agent + Agent Evaluation → ~$1,000/月 production cost
适合 Data Analyst?	Databricks SQL + AI/BI Dashboards + Metric Views + Alerts + Notebooks
适合 Data Scientist?	MLflow 3 + Model Registry + Model Serving + Feature Store + AutoML
适合 AI Agent 工程师?	Mosaic AI Foundation Models + Vector Search (= AI Search) + Custom Agents + Prompt Mgmt + Agent Eval

1. 2026 重大命名更新 (必看)

在你学之前,先记 3 个 2026 改名 — 旧文档可能用旧名:

旧名	新名	改名时间
Repos	Git folders	2026-03-11
Jobs	Lakeflow Jobs	2026-07-10
Vector Search	AI Search	2026-07-28 (项目重命名)

含义: 旧文档 / 旧 example 可能用旧名. 搜索 docs 用新名.

2. 12+ 服务速查表 (时间 + 难度)

#	服务	类别	部署时间	难度	30 秒 TL;DR
1	Databricks SQL	Data Access	5-10 min	⭐⭐	Serverless warehouse + Photon → 直接 query Delta Lake
2	AI/BI Dashboards	Data Access	30-45 min	⭐⭐⭐	Genie Code 自然语言生成 chart + filter cross-page
3	Metric Views	Data Access	45-60 min	⭐⭐⭐⭐	YAML 定义 net_revenue 跨 dashboard 复用
4	Alerts	Data Access	15-20 min	⭐⭐	Slack 集成 + Quartz cron + 4 个跨境支付 alert
5	MLflow 3 Tracking	ML Lifecycle	30 min	⭐⭐	Logged Models + 实验对比 + 跨境支付 fraud 完整代码
6	MLflow Model Registry	ML Lifecycle	40 min	⭐⭐⭐	Alias 治理 + 3 层 namespace + A/B test + 30 秒 rollback
7	Model Serving	ML Lifecycle	45 min	⭐⭐⭐	SDK/REST/UI 3 种 + Inference Table + 流量切分
8	Feature Store	ML Lifecycle	60 min	⭐⭐⭐⭐	Online Lakebase + Point-in-Time Join + Serving 自动 lookup
9	AutoML	ML Lifecycle	30 min	⭐⭐	Python API + 复用生成 code + 注册到 Registry
10	Mosaic AI Foundation Models	AI Agent	5-10 min	⭐	OpenAI-compatible API + DBRX/Llama 3.1/Mixtral + 3 种调用模式
11	Vector Search (= AI Search)	AI Agent	30-45 min	⭐⭐⭐	3 种 index + HNSW + RRF + SDK/UI 自动 embedding
12	Custom Agents	AI Agent	45-60 min	⭐⭐⭐⭐	LangGraph/LangChain 集成 + MLflow Tracing + production 部署
13	Prompt Management	AI Agent	15-20 min	⭐⭐	Version + manage prompts + 在 agent 里按版本调用
14	Agent Evaluation	AI Agent	30-45 min	⭐⭐⭐	LLM-as-a-judge + human feedback + MLflow 集成
15	Notebooks	共用	5 min	⭐	多语言 IDE + magic commands + dbutils + Git
16	Unity Catalog	共用	30 min	⭐⭐	3 层 namespace (catalog.schema.object) + 跨 data/model/feature
17	Git folders (原 Repos)	共用	20 min	⭐⭐	Git 集成 + clone + commit + PR workflow
18	Lakeflow Jobs (原 Jobs)	共用	15 min	⭐⭐	Cron schedule + retry + alert + cluster 自动启停

总部署时间: 假设你全做 (~10-12 小时 = 1 个工作日)

3. 4 大类 (12+ 服务) 难度排序

难度	服务数	学习时间	适合角色
⭐⭐ (Easy)	6	5-30 min	Data Analyst 入门
⭐⭐⭐ (Medium)	7	30-45 min	Data Scientist 熟悉
⭐⭐⭐⭐ (Hard)	3	45-60 min	Advanced ML/AI 工程师

4. 4 大常见组合 (per persona)

4.1 Data Analyst 组合 (5 服务, ~2.5 小时)

Databricks SQL + AI/BI Dashboards + Metric Views + Alerts + Notebooks
→ "SQL-based self-service analytics"

4.2 Data Scientist 组合 (5 服务, ~4 小时)

MLflow 3 + Model Registry + Model Serving + Feature Store + AutoML
→ "End-to-end ML lifecycle"

4.3 AI Agent 工程师组合 (5 服务, ~3.5 小时)

Mosaic AI Foundation Models + Vector Search + Custom Agents + Prompt Mgmt + Agent Eval
→ "Production agent observability"

4.4 共用 (4 服务, ~1.5 小时)

Notebooks + Unity Catalog + Git folders (Repos) + Lakeflow Jobs (Jobs)
→ "Cross-team collaboration + scheduling"

5. 关键洞察 (7 个)

5.1 Databricks SQL 是 5-10 分钟 — 多数服务最小可行部署

大多数企业用户从 SQL warehouse 起步. 5 分钟 setup 不夸张 (Community Edition).

5.2 MLflow 3 是 30-45 分钟 — Data Scientist 真正的入门门槛

需要懂 experiment tracking + Python. 但一旦会,所有 ML lifecycle 服务都基于此.

5.3 Mosaic AI Custom Agents 是 45-60 分钟 — 最高难度

需要懂 LangGraph / LangChain + MLflow Tracing. 跨境支付 agent 用例需要思考 prompt + tool design.

5.4 Notebooks + Unity Catalog 是任何角色必学 — 5-30 分钟

这 2 个是 platform 基础. 其他 14 个服务都基于这 2 个.

5.5 2026 命名更新重要

Repos → Git folders (Mar 11), Jobs → Lakeflow Jobs (Jul 10), Vector Search → AI Search (Jul 28). 旧文档 / 旧 example 可能用旧名.

5.6 真实踩坑 (vs 官方文档暗示)

• ⚠️ Chrome MCP navigation timeout (D+14 8-15 实证) — 用 curl fallback
• ⚠️ Databricks docs URL 经常 redirect → SPA 路由问题
• ⚠️ Firecrawl 配额耗尽 → 用 curl + HTMLParser
• ⚠️ Serverless SQL warehouse ≠ Classic cluster → 不要选错
• ⚠️ MLflow 3 Logged Models 概念没替代品 — 必须用

5.7 核心方法论: 8-14 vs 8-15 笔记互补

• 8-14 笔记: "是什么 / 现状 / 概述"
• 8-15 笔记: "怎么部署 / 多少时间 / 多少成本 / 哪里踩坑 / 跨境支付 use case"
• 8-14 + 8-15 互补 = 完整参考

6. 4 大类完整教程 (摘要)

6.1 Data Access 类 (Databricks SQL + AI/BI Dashboards + Metric Views + Alerts)

Databricks SQL — 5-10 分钟部署, ⭐⭐ 难度.

from databricks import sql

connection = sql.connect(
    server_hostname=".databricks.com",
    http_path="/sql/1.0/warehouses/abc123",
    access_token=""
)

cursor = connection.cursor()
cursor.execute("""
  SELECT DATE_TRUNC('day', pickup_datetime) AS day, COUNT(*) AS trips
  FROM samples.nyctaxi.trips
  GROUP BY 1
  ORDER BY 1
""")
results = cursor.fetchall()
print(results)

踩坑 1: 4 KB SQL 配额 (Community Edition) — 升级 Premium.

踩坑 2: Photon 关闭 (cluster 创建时手动选) — 失去 2-5x speedup.

踩坑 3: WC (Cluster) vs SE (Serverless) 类型混淆 — 用错 config.

踩坑 4: 跨 catalog 权限 miss — 报 "401 unauthorized". 用 GRANT 解决.

踩坑 5: Serverless SQL 太慢 — 用户并发 > 100, 升级到 Pro/Classic.

AI/BI Dashboards — 30-45 分钟, ⭐⭐⭐. Genie Code 自然语言生成 chart + filter cross-page.

Metric Views — 45-60 分钟, ⭐⭐⭐⭐. 解决"什么是 revenue" 跨部门争论.

Alerts — 15-20 分钟, ⭐⭐. 跨境支付 4 个 alert: 失败率 / 大额退款 / FX 波动 / slow query.

6.2 ML Lifecycle 类 (MLflow 3 + Model Registry + Model Serving + Feature Store + AutoML)

MLflow 3 Tracking — 30 分钟, ⭐⭐. Logged Models 跨整个 lifecycle track 一个 model.

import mlflow

mlflow.set_experiment("/Users/javis/payment_fraud")

with mlflow.start_run():
    mlflow.log_param("model_type", "lightgbm")
    mlflow.log_param("learning_rate", 0.05)

    model = lgb.train(params, train_data)

    mlflow.log_metric("auc", 0.92)
    mlflow.log_metric("precision", 0.85)

    mlflow.lightgbm.log_model(model, "model")
    mlflow.log_artifact("confusion_matrix.png")

MLflow Model Registry — 40 分钟, ⭐⭐⭐. Alias 治理 (champion / challenger) + 3 层 namespace + A/B test + 30 秒 rollback.

Model Serving — 45 分钟, ⭐⭐⭐. SDK/REST/UI 3 种创建 + Inference Table 自动记录 + 流量切分.

Feature Store — 60 分钟, ⭐⭐⭐⭐. 最难点 — Online Lakebase + Point-in-Time Join + Serving 自动 lookup.

AutoML — 30 分钟, ⭐⭐. Python API + 复用生成 code + 注册到 Registry.

6.3 AI Agent 类 (Mosaic AI Foundation Models + Vector Search + Custom Agents + Prompt Mgmt + Agent Eval)

Mosaic AI Foundation Models — 5-10 分钟, ⭐. 最简单的服务. OpenAI 兼容 client 即可调用 DBRX / Llama 3.1 / Mixtral / Claude / GPT.

from openai import OpenAI

client = OpenAI(
    api_key="",
    base_url="https://.databricks.com/serving-endpoints"
)

response = client.chat.completions.create(
    model="databricks-dbrx-instruct",
    messages=[
        {"role": "system", "content": "You are a payment support agent."},
        {"role": "user", "content": "How do I dispute a charge?"}
    ],
    max_tokens=200
)
print(response.choices[0].message.content)

3 种调用模式: Pay-per-token (default, 按 token) + Priority tier (低延迟, 高 cost) + Provisioned throughput (固定 latency, 固定 cost).

Vector Search (= AI Search 2026 Jul 28) — 30-45 分钟, ⭐⭐⭐. 3 种 index: Delta Sync managed (推荐) / Delta Sync self-managed / Direct Vector Access. HNSW + RRF.

Custom Agents — 45-60 分钟, ⭐⭐⭐⭐ 最高难度. LangGraph / LangChain / OpenAI 集成 + MLflow Tracing 自动 + production 部署.

Prompt Management — 15-20 分钟, ⭐⭐. Version + manage prompts + A/B test + 30 秒 rollback.

Agent Evaluation — 30-45 分钟, ⭐⭐⭐. LLM-as-a-judge + human feedback + MLflow 集成.

6.4 共用类 (Notebooks + Unity Catalog + Git folders + Lakeflow Jobs)

Notebooks — 5 分钟, ⭐. 多语言 IDE + magic commands + dbutils + widgets + Git.

# MAGIC %sql
# MAGIC SELECT * FROM samples.nyctaxi.trips LIMIT 10;

result = spark.sql("SELECT COUNT(*) FROM main.default.payments")
display(result)

Unity Catalog — 30 分钟, ⭐⭐. 3 层 namespace (catalog.schema.object) 跨 data + model + feature 统一治理.

Git folders (原 Repos, Mar 11 改名) — 20 分钟, ⭐⭐. Git 集成 + clone + commit + PR workflow.

Lakeflow Jobs (原 Jobs, Jul 10 改名) — 15 分钟, ⭐⭐. Cron schedule + retry + alert + cluster 自动启停. 2026 改名 — search 时用 "Lakeflow Jobs" 不要 "Jobs".

7. 完整 RAG + 跨境支付 Agent 端到端 (5 服务串联)

这是 §6.3 AI Agent 5 服务的完整串联示例 — 跨境支付 (Payoneer) 客服 RAG + agent 完整代码:

# 1. Vector Search (AI Search) — 检索 FAQ
from databricks.vector_search.client import VectorSearchClient

client = VectorSearchClient()

def retrieve_faq(question, top_k=3):
    results = client.similarity_search(
        index_name="main.finance.payment_faq_index",
        query_text=question,
        num_results=top_k
    )
    return "\n".join([r["faq_answer"] for r in results["result"]["data_array"]])

# 2. Foundation Model + Prompt Mgmt — LLM 调用 with versioned prompt
from openai import OpenAI
import mlflow

client = OpenAI(api_key=..., base_url="https://payoneer.datababricks.com/serving-endpoints")

def query_llm(question, faq_context, prompt_version="latest"):
    prompt = mlflow.load_prompt(
        f"prompts:/payment_support_agent_prompt@{prompt_version}"
    )
    response = client.chat.completions.create(
        model="databricks-dbrx-instruct",
        messages=[
            {"role": "system", "content": prompt.format(context=faq_context)},
            {"role": "user", "content": question}
        ],
        max_tokens=300,
        temperature=0.3
    )
    return response.choices[0].message.content

# 3. Custom Agent — Agent with tool calling
class PaymentSupportAgent(mlflow.pyfunc.ResponsesAgent):
    @mlflow.trace
    def predict(self, context, messages):
        question = messages[-1]["content"]

        # 3.1 RAG retrieval
        faq_context = retrieve_faq(question)

        # 3.2 LLM call
        response = query_llm(question, faq_context)

        # 3.3 Tool call if refund request
        if "refund" in question.lower():
            response += f"\n\n[系统追踪] 退款已启动..."

        return response

# 4. MLflow Tracing 自动 (Service 3 内置)
# Production 部署后, MLflow UI 自动显示 agent trace

# 5. Agent Evaluation — 自动评估
results = mlflow.evaluate(
    model="models:/payment_support_agent/Production",
    data=eval_data,  # 20 个 test cases
    evaluators=["faithfulness", "relevance", "safety"]
)
print(results.metrics)
# {'faithfulness/v1/mean': 4.2, 'relevance/v1/mean': 4.5, ...}

完整 workflow: Vector Search → Foundation Model + Prompt Mgmt → Custom Agent → MLflow Tracing 自动 → Agent Evaluation 自动.

每秒成本: ~$0.005 (DBRX + Vector Search + MLflow tracing). 1M calls/月 ≈ $1,000/月.

8. 成本估算 (按 2026 Databricks 公开价格)

模式	月费	适合
Community Edition	$0	学习 / POC
Free Trial	$0 (14 天)	短期 pilot
Serverless PAYG	$0.22/DBU + AWS	小团队 / 试点
Premium (PAYG)	$0.55/DBU + AWS	标准 production
Enterprise	协商	大型 / 合规

新手月成本估算 (5 TB data + 100 queries/day + 10 ML runs):

• 1 Serverless SQL warehouse + 1 Notebooks: ~$300-500/mo
• 加 MLflow 3 + Model Serving: ~$800-1500/mo
• 加 Mosaic AI Vector Search + Custom Agents: ~$1500-3000/mo

9. 实盘 Payoneer 跨境支付 cost 试算

场景: Payoneer 1M 笔/月 跨境支付客服 (5 服务 RAG + agent):

服务	Usage	Cost
Foundation Model (DBRX)	500 tokens × 1M = 500M tokens	$700/月
Vector Search (AI Search)	100K FAQ + 1M queries	$10/月
Custom Agent (Model Serving)	Small endpoint + 5 iterations	$120/月
Prompt Management	5 versions	$0
Agent Evaluation	5 evals/month × 20 cases	$50/eval
Total		~$880/月 + $50/eval

对比 Claude 3.5 Sonnet (质量更高): $4,500/月 — DBRX 节省 80%.

10. D+10 实战选择 (Javis 视角)

当前: Data Analyst @ Payoneer (BigQuery)

Week 1 path (最低 priority):

• Databricks SQL + AI/BI Dashboards + Alerts (= Data Analyst 组合)
• 评估 Databricks SQL vs BigQuery 性能 + cost

D+10 target: Data Scientist

Week 2-3 path:

• MLflow 3 + Model Registry + Model Serving + Feature Store
• 选 1 个跨境支付 fraud detection 实战项目

D+10 interest: AI Agent Engineer

Week 4-5 path:

• Mosaic AI Foundation Models + Vector Search + Custom Agents + Agent Evaluation
• 选 1 个 RAG / agent 跨境支付场景

共用基础设施 (任何时候)

• Notebooks + Unity Catalog + Git folders + Lakeflow Jobs

总 D+10 实战周期: 4-5 周 → 完整覆盖 12+ 服务.

11. 5 大实战踩坑 (D+14 8-15 实证)

1. Chrome MCP navigation timeout — Databricks docs 经常 timeout. fallback: curl + Python HTMLParser.
2. Databricks docs URL 经常 redirect — SPA 路由. /agents/ + /ai-search/ 是 2026 新路径.
3. Firecrawl 配额耗尽 — web_search / web_extract 用不了. fallback: curl + Python HTMLParser.
4. Serverless SQL ≠ Classic cluster — 不要选错. 你需要的是 Serverless SQL warehouse 不是 Classic cluster.
5. MLflow 3 Logged Models 概念没替代品 — 必须用. 别退回 mlflow.log_model() 上古.

12. 完整 5 篇笔记 + 8-15 Vault Cross-references

本文是 5 篇 D+14 8-15 教程的综合索引. 完整 step-by-step 教程请看:

• 01 - Data Access Services (50 KB / 1321 行) — Databricks SQL + AI/BI Dashboards + Metric Views + Alerts
• 02 - ML Lifecycle Services (82 KB / 2055 行) — MLflow 3 + Model Registry + Model Serving + Feature Store + AutoML
• 03 - AI Agent Services (40 KB / 1141 行) — Mosaic AI Foundation Models + Vector Search + Custom Agents + Prompt Mgmt + Agent Eval
• 04 - Common Services (105 KB / 2905 行) — Notebooks + Unity Catalog + Git folders + Lakeflow Jobs
• 00 - Index + 12 Services Quick Reference (11 KB) — 速查表 + 4 大类组合

合计 295 KB / 7422 行.

13. 参考资源

官方文档 (2026-08-05 最新)

• Databricks SQL · 2026-07-17
• MLflow on Databricks · 2026-08-05
• Foundation Model APIs · 2026-07-28
• AI Search (formerly Vector Search) · 2026-07-28
• Agents Overview · 2026-08-03
• Custom Agents · 2026-08-03
• Databricks Notebooks · 2026-06-16
• Repos (now Git folders) · 2026-03-11
• Jobs (now Lakeflow Jobs) · 2026-07-10

你 Vault 里的相关

• Databricks Spark Architecture + Deployment · 集群部署
• Databricks vs Snowflake vs BigQuery Comparison · 3 家对比
• Databricks for Data Scientists and Analysts · 12+ 服务 overview

---