mirror of
https://github.com/MariaDB/server.git
synced 2026-10-23 18:49:29 +03:00
96aee50cfb
Problem: ======== OPTIMIZE TABLE on a FULLTEXT index crashes when the aux tables hold two spellings of a word that are equal under the index collation. The index lookup in fts_index_fetch_nodes() and the DELETE in fts_optimize_write_word() both compare words by collation, but fts_optimize_index_fetch_node() compares the bytes, so each spelling becomes a separate fts_word_t. fts_optimize_compact() then optimizes and writes back each of them separately. Writing one deletes the rows of the other, including the rows that were just written, so index entries can be lost. If all the documents were deleted, the first write removes both words. The next word that fts_optimize_words() reads from the zip buffer is the other spelling, the fetch finds nothing, and the server aborts Solution: ======== fts_optimize_collation_fetch_node(): New callback used only by OPTIMIZE. It adds the fetched records whose words are equal under the index collation to the same fts_word_t, so the spellings are optimized and written back as one word. The records arrive ordered by first_doc_id, so the nodes stay in order. fts_optimize_fetch_node(): The helper of the fetch callbacks, which compares the words either by bytes or by the given collation. fts_optimize_index_fetch_node() still compares by bytes, which is what the information_schema code expects. fts_optimize_words(): Use fts_optimize_collation_fetch_node(). The word read from the zip buffer can already have been optimized together with an earlier spelling, so an empty fetch is not an error. Skip the word instead of asserting.