From: Vladimir Davydov <vdavydov.dev@gmail.com> To: kostja@tarantool.org Cc: tarantool-patches@freelists.org Subject: [PATCH 02/25] vinyl: simplify vy_squash_process Date: Fri, 27 Jul 2018 14:29:42 +0300 [thread overview] Message-ID: <00e612f6de645f182d378c00d7ddb77d0c26b303.1532689066.git.vdavydov.dev@gmail.com> (raw) In-Reply-To: <cover.1532689065.git.vdavydov.dev@gmail.com> In-Reply-To: <cover.1532689065.git.vdavydov.dev@gmail.com> Since vy_point_lookup() now guarantees that it returns the newest tuple version, we can remove the code that squashes UPSERTs from vy_squash_process(). --- src/box/vinyl.c | 115 ++++++-------------------------------------------------- 1 file changed, 12 insertions(+), 103 deletions(-) diff --git a/src/box/vinyl.c b/src/box/vinyl.c index 374c5252..530820e8 100644 --- a/src/box/vinyl.c +++ b/src/box/vinyl.c @@ -3532,11 +3532,6 @@ vy_squash_process(struct vy_squash *squash) struct vy_lsm *lsm = squash->lsm; struct vy_env *env = squash->env; - /* - * vy_apply_upsert() is used for primary key only, - * so this is the same as lsm->key_def - */ - struct key_def *def = lsm->cmp_def; /* Upserts enabled only in the primary index LSM tree. */ assert(lsm->index_id == 0); @@ -3554,8 +3549,10 @@ vy_squash_process(struct vy_squash *squash) /* * While we were reading on-disk runs, new statements could - * have been inserted into the in-memory tree. Apply them to - * the result. + * have been prepared for the squashed key. We mustn't apply + * them, because they may be rolled back, but we must adjust + * their n_upserts counter so that they will get squashed by + * vy_lsm_commit_upsert(). */ struct vy_mem *mem = lsm->mem; struct tree_mem_key tree_key = { @@ -3572,108 +3569,20 @@ vy_squash_process(struct vy_squash *squash) tuple_unref(result); return 0; } - /** - * Algorithm of the squashing. - * Assume, during building the non-UPSERT statement - * 'result' in the mem some new UPSERTs were inserted, and - * some of them were commited, while the other were just - * prepared. And lets UPSERT_THRESHOLD to be equal to 3, - * for example. - * Mem - * -------------------------------------+ - * UPSERT, lsn = 1, n_ups = 0 | - * UPSERT, lsn = 2, n_ups = 1 | Commited - * UPSERT, lsn = 3, n_ups = 2 | - * -------------------------------------+ - * UPSERT, lsn = MAX, n_ups = 3 | - * UPSERT, lsn = MAX + 1, n_ups = 4 | Prepared - * UPSERT, lsn = MAX + 2, n_ups = 5 | - * -------------------------------------+ - * In such a case the UPSERT statements with - * lsns = {1, 2, 3} are squashed. But now the n_upsert - * values in the prepared statements are not correct. - * If we will not update values, then the - * vy_lsm_commit_upsert will not be able to squash them. - * - * So after squashing it is necessary to update n_upsert - * value in the prepared statements: - * Mem - * -------------------------------------+ - * UPSERT, lsn = 1, n_ups = 0 | - * UPSERT, lsn = 2, n_ups = 1 | Commited - * REPLACE, lsn = 3 | - * -------------------------------------+ - * UPSERT, lsn = MAX, n_ups = 0 !!! | - * UPSERT, lsn = MAX + 1, n_ups = 1 !!! | Prepared - * UPSERT, lsn = MAX + 2, n_ups = 2 !!! | - * -------------------------------------+ - */ vy_mem_tree_iterator_prev(&mem->tree, &mem_itr); - const struct tuple *mem_stmt; - int64_t stmt_lsn; - /* - * According to the described algorithm, squash the - * commited UPSERTs at first. - */ + uint8_t n_upserts = 0; while (!vy_mem_tree_iterator_is_invalid(&mem_itr)) { + const struct tuple *mem_stmt; mem_stmt = *vy_mem_tree_iterator_get_elem(&mem->tree, &mem_itr); - stmt_lsn = vy_stmt_lsn(mem_stmt); - if (vy_tuple_compare(result, mem_stmt, def) != 0) - break; - /** - * Leave alone prepared statements; they will be handled - * in vy_range_commit_stmt. - */ - if (stmt_lsn >= MAX_LSN) + if (vy_tuple_compare(result, mem_stmt, lsm->cmp_def) != 0 || + vy_stmt_type(mem_stmt) != IPROTO_UPSERT) break; - if (vy_stmt_type(mem_stmt) != IPROTO_UPSERT) { - /** - * Somebody inserted non-upsert statement, - * squashing is useless. - */ - tuple_unref(result); - return 0; - } - assert(lsm->index_id == 0); - struct tuple *applied = vy_apply_upsert(mem_stmt, result, def, - mem->format, true); - lsm->stat.upsert.applied++; - tuple_unref(result); - if (applied == NULL) - return -1; - result = applied; - /** - * In normal cases we get a result with the same lsn as - * in mem_stmt. - * But if there are buggy upserts that do wrong things, - * they are ignored and the result has lower lsn. - * We should fix the lsn in any case to replace - * exactly mem_stmt in general and the buggy upsert - * in particular. - */ - vy_stmt_set_lsn(result, stmt_lsn); + assert(vy_stmt_lsn(mem_stmt) >= MAX_LSN); + vy_stmt_set_n_upserts((struct tuple *)mem_stmt, n_upserts); + if (n_upserts <= VY_UPSERT_THRESHOLD) + ++n_upserts; vy_mem_tree_iterator_prev(&mem->tree, &mem_itr); } - /* - * The second step of the algorithm above is updating of - * n_upsert values of the prepared UPSERTs. - */ - if (stmt_lsn >= MAX_LSN) { - uint8_t n_upserts = 0; - while (!vy_mem_tree_iterator_is_invalid(&mem_itr)) { - mem_stmt = *vy_mem_tree_iterator_get_elem(&mem->tree, - &mem_itr); - if (vy_tuple_compare(result, mem_stmt, def) != 0 || - vy_stmt_type(mem_stmt) != IPROTO_UPSERT) - break; - assert(vy_stmt_lsn(mem_stmt) >= MAX_LSN); - vy_stmt_set_n_upserts((struct tuple *)mem_stmt, - n_upserts); - if (n_upserts <= VY_UPSERT_THRESHOLD) - ++n_upserts; - vy_mem_tree_iterator_prev(&mem->tree, &mem_itr); - } - } lsm->stat.upsert.squashed++; -- 2.11.0
next prev parent reply other threads:[~2018-07-27 11:29 UTC|newest] Thread overview: 39+ messages / expand[flat|nested] mbox.gz Atom feed top 2018-07-27 11:29 [PATCH 00/25] vinyl: eliminate disk read on REPLACE/DELETE Vladimir Davydov 2018-07-27 11:29 ` [PATCH 01/25] vinyl: make point lookup always return the latest tuple version Vladimir Davydov 2018-07-27 11:29 ` Vladimir Davydov [this message] 2018-07-27 11:29 ` [PATCH 03/25] vinyl: always get full tuple from pk after reading from secondary index Vladimir Davydov 2018-07-27 11:29 ` [PATCH 04/25] vinyl: fold vy_replace_one and vy_replace_impl Vladimir Davydov 2018-07-27 11:29 ` [PATCH 05/25] vinyl: fold vy_delete_impl Vladimir Davydov 2018-07-27 11:29 ` [PATCH 06/25] vinyl: refactor unique check Vladimir Davydov 2018-07-27 11:29 ` [PATCH 07/25] vinyl: check key uniqueness before modifying tx write set Vladimir Davydov 2018-07-27 11:29 ` [PATCH 08/25] vinyl: remove env argument of vy_check_is_unique_{primary,secondary} Vladimir Davydov 2018-07-31 20:45 ` [tarantool-patches] " Konstantin Osipov 2018-07-27 11:29 ` [PATCH 09/25] vinyl: store full tuples in secondary index cache Vladimir Davydov 2018-07-31 20:47 ` Konstantin Osipov 2018-07-27 11:29 ` [PATCH 10/25] vinyl: do not free pending tasks on shutdown Vladimir Davydov 2018-07-31 20:48 ` Konstantin Osipov 2018-07-27 11:29 ` [PATCH 11/25] vinyl: store pointer to scheduler in struct vy_task Vladimir Davydov 2018-07-31 20:49 ` Konstantin Osipov 2018-07-27 11:29 ` [PATCH 12/25] vinyl: rename some members of vy_scheduler and vy_task struct Vladimir Davydov 2018-07-27 11:29 ` [PATCH 13/25] vinyl: use cbus for communication between scheduler and worker threads Vladimir Davydov 2018-07-27 11:29 ` [PATCH 14/25] vinyl: zap vy_scheduler::is_worker_pool_running Vladimir Davydov 2018-07-27 11:29 ` [PATCH 15/25] vinyl: rename vy_task::status to is_failed Vladimir Davydov 2018-07-27 11:29 ` [PATCH 16/25] xrow: allow to store flags in DML requests Vladimir Davydov 2018-07-27 11:29 ` [PATCH 17/25] vinyl: pin last statement returned by write iterator explicitly Vladimir Davydov 2018-07-27 11:29 ` [PATCH 18/25] vinyl: teach write iterator to return overwritten tuples Vladimir Davydov 2018-07-27 11:29 ` [PATCH 19/25] vinyl: prepare write iterator heap comparator for deferred DELETEs Vladimir Davydov 2018-07-27 11:30 ` [PATCH 20/25] vinyl: allow to skip certain statements on read Vladimir Davydov 2018-07-27 11:30 ` [PATCH 21/25] vinyl: add function to create surrogate deletes from raw msgpack Vladimir Davydov 2018-07-27 11:30 ` [PATCH 22/25] vinyl: remove pointless assertion from vy_stmt_new_surrogate_delete Vladimir Davydov 2018-07-27 11:30 ` [PATCH 23/25] txn: add helper to detect transaction boundaries Vladimir Davydov 2018-07-31 20:52 ` [tarantool-patches] " Konstantin Osipov 2018-07-27 11:30 ` [PATCH 24/25] Introduce _vinyl_deferred_delete system space Vladimir Davydov 2018-07-31 20:54 ` Konstantin Osipov 2018-08-01 14:00 ` Vladimir Davydov 2018-08-01 20:25 ` [tarantool-patches] " Konstantin Osipov 2018-08-02 9:43 ` Vladimir Davydov 2018-08-06 8:42 ` Vladimir Davydov 2018-07-27 11:30 ` [PATCH 25/25] vinyl: eliminate disk read on REPLACE/DELETE Vladimir Davydov 2018-07-31 20:55 ` Konstantin Osipov 2018-08-01 16:03 ` Vladimir Davydov 2018-08-01 16:51 ` Vladimir Davydov
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=00e612f6de645f182d378c00d7ddb77d0c26b303.1532689066.git.vdavydov.dev@gmail.com \ --to=vdavydov.dev@gmail.com \ --cc=kostja@tarantool.org \ --cc=tarantool-patches@freelists.org \ --subject='Re: [PATCH 02/25] vinyl: simplify vy_squash_process' \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: link
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox