为什么子查询中的 OR 会使查询慢得多？

Question

我正在使用 MySQL，并且有以下我正在尝试改进的查询：

SELECT
    *
FROM
    overpayments AS op
    JOIN payment_allocations AS overpayment_pa ON overpayment_pa.allocatable_id = op.id
    AND overpayment_pa.allocatable_type = 'Overpayment'
    JOIN (
        SELECT
            pa.payment_source_type,
            pa.payment_source_id,
            ft.conversion_rate
        FROM
            payment_allocations AS pa
            LEFT JOIN line_items AS li ON pa.payment_source_id = li.id
            LEFT JOIN credit_notes AS cn ON li.parent_document_id = cn.id
            LEFT JOIN financial_transactions AS ft ON (
                ft.commercial_document_id = pa.payment_source_id
                AND ft.commercial_document_type = pa.payment_source_type
            )
            OR (
                ft.commercial_document_id = cn.id
                AND ft.commercial_document_type = 'CreditNote'
            )
        WHERE
            pa.allocatable_type = 'Overpayment'
            AND pa.company_id = 14792
            AND ft.company_id = 14792
    ) AS op_bank_transaction_ft ON op_bank_transaction_ft.payment_source_id = overpayment_pa.payment_source_id
    AND op_bank_transaction_ft.payment_source_type = overpayment_pa.payment_source_type;

运行需要 10 秒。通过删除子查询中的 OR 语句并使用 COALESCE 获得结果，我能够将其提高到 0.047 秒：

SELECT
    *
FROM
    overpayments AS op
    JOIN payment_allocations AS overpayment_pa ON overpayment_pa.allocatable_id = op.id
    AND overpayment_pa.allocatable_type = 'Overpayment'
    JOIN (
        SELECT
            pa.payment_source_type,
            pa.payment_source_id,
            coalesce(ft_one.conversion_rate, ft_two.conversion_rate)
        FROM
            payment_allocations AS pa
            LEFT JOIN line_items AS li ON pa.payment_source_id = li.id
            LEFT JOIN credit_notes AS cn ON li.parent_document_id = cn.id
            LEFT JOIN financial_transactions AS ft_one ON (
                ft_one.commercial_document_id = pa.payment_source_id
                AND ft_one.commercial_document_type = pa.payment_source_type
                AND ft_one.company_id = 14792
            )
            LEFT JOIN financial_transactions AS ft_two ON (
                ft_two.commercial_document_id = cn.id
                AND ft_two.commercial_document_type = 'CreditNote'
                AND ft_two.company_id = 14792
            )
        WHERE
            pa.allocatable_type = 'Overpayment'
            AND pa.company_id = 14792
            
    ) AS op_bank_transaction_ft ON op_bank_transaction_ft.payment_source_id = overpayment_pa.payment_source_id
    AND op_bank_transaction_ft.payment_source_type = overpayment_pa.payment_source_type;

但是，我真的不明白为什么会这样？原来的子查询跑得很快，只返回了 2 个结果，那为什么它会减慢查询速度呢？解释第一个查询返回以下内容：

# id	select_type	table	partitions	type	possible_keys	key	key_len	ref	rows	filtered	Extra	FIELD13
1	SIMPLE	pa		ref	index_payment_allocations_on_payment_source_id	index_payment_allocations_on_company_id	index_payment_allocations_on_company_id	5	const	191	10.00	Using where
1	SIMPLE	overpayment_pa		ref	index_payment_allocations_on_payment_source_id	index_payment_allocations_on_allocatable_id	index_payment_allocations_on_payment_source_id	5	rails.pa.payment_source_id	1	3.42	Using where
1	SIMPLE	op		eq_ref	PRIMARY	PRIMARY	4	rails.overpayment_pa.allocatable_id	1	100.00
1	SIMPLE	li		eq_ref	PRIMARY	PRIMARY	4	rails.pa.payment_source_id	1	100.00
1	SIMPLE	cn		eq_ref	PRIMARY	PRIMARY	8	rails.li.parent_document_id	1	100.00	Using where; Using index
1	SIMPLE	ft		ALL	transactions_unique_by_commercial_doc				12587878	0.00	Range checked for each record (index map: 0x2)

第二次我得到以下信息：

# id	select_type	table	partitions	type	possible_keys	key	key_len	ref	rows	filtered	Extra	FIELD13	FIELD14
1	SIMPLE	pa		ref	index_payment_allocations_on_payment_source_id	index_payment_allocations_on_company_id	index_payment_allocations_on_company_id	5	const	191	10.00	Using where
1	SIMPLE	overpayment_pa		ref	index_payment_allocations_on_payment_source_id	index_payment_allocations_on_allocatable_id	index_payment_allocations_on_payment_source_id	5	rails.pa.payment_source_id	1	3.42	Using where
1	SIMPLE	op		eq_ref	PRIMARY	PRIMARY	4	rails.overpayment_pa.allocatable_id	1	100.00
1	SIMPLE	ft_one		ref	transactions_unique_by_commercial_doc	index_financial_transactions_on_company_id	transactions_unique_by_commercial_doc	773	rails.pa.payment_source_id	rails.pa.payment_source_type	1	100.00	Using where
1	SIMPLE	li		eq_ref	PRIMARY	PRIMARY	4	rails.pa.payment_source_id	1	100.00
1	SIMPLE	cn		eq_ref	PRIMARY	PRIMARY	8	rails.li.parent_document_id	1	100.00	Using where; Using index
1	SIMPLE	ft_two		ref	transactions_unique_by_commercial_doc	index_financial_transactions_on_company_id	transactions_unique_by_commercial_doc	773	rails.cn.id	const	1	100.00	Using where

但我真的不知道如何解释这些结果。

Answer 1

查看您的第一个 EXPLAIN 的最后一行的右侧。它不使用索引，它必须扫描百万行。那很慢。您的第二个查询对查询的每一步都使用了索引，因此速度要快得多。

如果您的第二个查询产生了正确的结果，请使用它并且不要回头。恭喜！您已优化查询。

OR 操作，尤其是在 ON 子句中，比通常的查询计划器模块更难满足，因为它们通常意味着它必须采用两个单独的子查询的并集。看起来计划者选择在你的情况下强行使用它。 （蛮力 === 扫描多行。）

不知道您的索引，很难进一步帮助您。

阅读本文以了解更多信息。 https://use-the-index-luke.com

Answer 2

这些可能会进一步加快第二个公式：

overpayment_pa:
    INDEX(payment_source_id, payment_source_type, allocatable_type, allocatable_id)
pa: INDEX(allocatable_type, company_id, payment_source_id,  payment_source_type)
financial_transactions:
    INDEX(commercial_document_id, commercial_document_type, company_id, conversion_rate)