为什么 Postgres 11 哈希索引这么大？

Question

RDS 上的 Postgres 11.4 和家里的 11.5。

我今天更仔细地查看哈希索引，因为我遇到了 citext 索引被忽略的问题。而且我发现我不明白为什么哈希索引如此之大。当我预计它需要 10 个字节 + 一些开销时，它需要大约 50 个字节/行。

我有一个示例数据库，其中包含一个名为 record_changes_log_detail 的表，该表有 7,733,552 条记录，因此约为 8M。在该表中有一个名为 old_value 的 citext 字段，它是哈希索引的来源：

CREATE INDEX record_changes_log_detail_old_value_ix_hash
    ON record_changes_log_detail
    USING hash (old_value);

这是对索引大小的检查：

select
'record_changes_log_detail_old_value_ix_hash' as index_name,
pg_relation_size ('record_changes_log_detail_old_value_ix_hash') as bytes,
pg_size_pretty(pg_relation_size ('record_changes_log_detail_old_value_ix_hash')) as pretty

这将返回 379,322,368 字节，或大约 362MB。我已经深入挖掘了源代码，并且对这篇精美的作品进行了更多研究。

听起来像行的哈希索引条目是与哈希键本身配对的 TID。以及页面内的某种索引计数器。那是两个 4 字节的整数，我猜是 1 或 2 字节的整数。作为一个简单的计算，10 字节 * 7,733,552 = 77,335,520。实际索引大约比那个大 5 倍。诚然，您需要为索引结构本身提供空间，但它不应该将每行的粗略成本从 ~10 字节到 ~50 字节，对吗？

这是索引的详细信息，使用pageinspect扩展阅读，然后手动旋转以提高可读性。

select * 
from hash_metapage_info(get_raw_page('record_changes_log_detail_old_value_ix_hash',0));


magic   105121344
version 4
ntuples 7733552
ffactor 307
bsize   8152
bmsize  4096
bmshift 15
maxbucket   28671
highmask    32767
lowmask 16383
ovflpoint   32
firstfree   17631
nmaps   1
procid  17269
spares  {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,17631,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0}
mapp    {28673,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0}

select *
from hash_page_stats(get_raw_page('record_changes_log_detail_old_value_ix_hash',1));

live_items  2
dead_items  0
page_size   8192
free_size   8108
hasho_prevblkno 28671
hasho_nextblkno 4294967295
hasho_bucket    0
hasho_flag  2
hasho_page_id   65408

Answer 1

我得到了一个新构建的索引，大小为 256MB，表大小也相同。您的索引是新建立的吗？是否在构建之前对表进行了新分析（索引是根据表中的估计行预先确定大小的）。您的重复项分布情况如何？

事物以最小 8 字节对齐方式存储，因此哈希索引元组是 16 字节，即使它应该适合 10（或 12，或其他）。哈希页平均只有一半满。桶按预定顺序分裂，它必须分裂下一个桶，而不是最满的桶。

select *
from hash_page_stats(get_raw_page('record_changes_log_detail_old_value_ix_hash',1));

live_items  2
dead_items  0
page_size   8192
free_size   8108

仅查看一页您不会学到很多东西，但该页奇怪地缺乏元组。也许你有一个病态的数据分布。

对数据库进行微观管理到这种程度是不值得的。