第 22 章 四通八达：服务接入、连接池与路由

把这笔短事务交给当前可写主库
把这批可容忍陈旧的查询交给合格副本
让迁移工具跟随主库，但不要经过事务池
把重型分析限制在指定离线副本

role          选择主库、副本还是指定成员
consistency   允许多旧，是否要求 read-your-writes
path          经过代理、连接池还是直达 PostgreSQL
session       客户端会话是否稳定绑定一个 backend
capacity      最多建立多少 client/server connection，在哪里排队
failure       旧连接如何结束，新连接何时恢复，提交结果如何判定
security      使用哪个身份、HBA/TLS/证书与审计边界
discovery     一个地址、VIP、DNS、多 host 还是驱动拓扑发现

业务动作
  -> 一致性和会话要求
      -> 语义服务端点
          -> PostgreSQL backend 成本
              -> 全局连接与并发预算
                  -> PgBouncer 模式及兼容性
                      -> HAProxy/Patroni 角色路由
                          -> 断连、重试与结果判定
                              -> Pigsty 声明和渲染
                                  -> 端到端演练与生产差距

Patroni 有 leader
  != 应用主写端点可用

HAProxy backend UP
  != PgBouncer 身份、会话语义正确

SHOW POOLS 有空位
  != 业务延迟和数据库并发安全

连接失败后重试成功
  != 上一次写入一定没有提交

target          pg36-l2-vagrant/pg-test
entry           10.10.10.11 HAProxy
Pigsty          v4.4.0
PostgreSQL      18.4
PgBouncer       1.25.2, transaction mode
HAProxy         3.4.2
fixture         database/user test/test, schema pg36_ch22
initial         pg-test-1 primary, timeline 9
forward         pg-test-2 primary, timeline 10
restored        pg-test-1 primary, timeline 11

default_pool_size=50
reserve_pool_size=30
query_wait_timeout=120

default_pool_size=2
reserve_pool_size=0
query_wait_timeout=5

endpoint semantics                      4 / 4 matched
concurrent clients                      12
maximum active PostgreSQL backends       2
maximum waiting clients                 10
fastest / slowest query              254 / 1519 ms

session SET stayed with backend          yes
same state leaked to another client      yes
original client got another backend      yes

protocol prepared iterations            12 / 12 correct
backend reassignment                     observed
SQL PREPARE after reassignment           SQLSTATE 26000

replica token visibility                 11.092 ms
interpretation                           one observation only

forward command                          2.774 s
restore command                          2.766 s
acknowledged writes                      339
unknown outcomes                         48, all absent after lookup
acknowledged missing                       0
duplicate token                            0
unreconciled unknown                       0
maximum conservative write gap          8.510 s
counterexamples rejected                  15

four service paths on one entry          通过
two-slot queue and backpressure          通过
transaction-pool session hazard          通过
protocol prepared exact version pair     通过
SQL PREPARE cross-backend failure        通过
one async visibility observation         通过
two healthy planned switchovers          通过
token reconciliation                     通过
original pool settings/final leader       恢复

redundant HAProxy/VIP/DNS entry           未证明
client/server TLS                         未验收
representative production load           未测试
driver/ORM version matrix                 未测试
unplanned failover/fencing                未测试
bounded replica staleness                 未承诺
production approval                       pending

EX22-SINGLE-HAPROXY-ENTRY
  只从 10.10.10.11 进入；不能声称入口冗余。

EX22-ASYNC-READ-OBSERVATION
  只采样一个 token；不能声称 read-your-writes 或 bounded staleness。

EX22-SYNTHETIC-LOAD
  负载小且运行在 laptop sandbox；不能推出生产容量。

EX22-RUNTIME-POOL-OVERRIDE
  只临时改变一个 PgBouncer 进程并恢复；不是声明式生产策略验收。