Hi! Hope you did it with piña colada under the palm tree;-) Regarding MDEV-31646 preserve DMLs in case of online binlog fault: It's changes are shown there as a part of ca64ddcc709, but it's important to point out that it's a separate thing: It is the commit that introduces error reporting to the ALTER thread, and its practically important only for a 32-bit build. I made it, but not sure if it's worth it. It complicates the architecture and brings many ad-hoc checks into the code. What do you think? Though if some multi-writer IO_CACHE replacement will finally be deployed, it may remove it all, since the 4GB restriction will not take place anymore (the transaction cache will still be an issue though). As an alternative, we could just fix IO_CACHE to support files larger than SIZE_T_MAX, by making end_of_file ulonglong, and all the corresponding changes. On Thu, 20 Jul 2023 at 17:59, Sergei Golubchik <serg@mariadb.org> wrote:
Hi, Nikita,
This is a review of all the three commits
On Jul 20, Nikita Malyavin wrote:
revision-id: ca64ddcc709 (mariadb-11.0.1-144-gca64ddcc709) parent(s): f34c7419cb8 author: Nikita Malyavin committer: Nikita Malyavin timestamp: 2023-07-19 02:31:32 +0400 message:
MDEV-31646 Online alter applies binlog cache limit to cache writes
diff --git a/sql/log.cc b/sql/log.cc index afd0643fe82..7466d63d325 100644 --- a/sql/log.cc +++ b/sql/log.cc @@ -2291,8 +2294,14 @@ int binlog_log_row_online_alter(TABLE* table, const uchar *before_record, before_record, after_record);
table->rpl_write_set= old_rpl_write_set; + thd->pop_internal_handler(); + + if (unlikely(error)) + push_warning_printf(thd, Sql_condition::WARN_LEVEL_NOTE, ER_DISK_FULL, + "Broken online alter log. " + "ALTER TABLE will finish with error.");
I don't think so. Why would a DML statement get a warning about the concurrently running ALTER TABLE? It didn't do anything wrong, I don't think there should be a warning here
Like, someone is running a series of DML statements, or the same statement many times. At some point another connection runs ALTER TABLE and suddenly those perfectly good DML statements start spewing out warnings. ALTER ends and all is clear again - looks wrong to me
Honestly I added this printf only to pass buildbot (unused `error`), and I don't want to leave the result unhandled. But you see, if for example server has binlog enabled, this perfect DML will simply fail! And this warning is about that some system limit is reached.
- return unlikely(error) ? HA_ERR_RBR_LOGGING_FAILED : 0; + return 0; }
static void @@ -2592,7 +2601,7 @@ void Event_log::set_write_error(THD *thd, bool is_transactional) DBUG_VOID_RETURN; }
-bool Event_log::check_write_error(THD *thd) +bool MYSQL_BIN_LOG::check_write_error(THD *thd)
What's the difference? it's a static method anyway, FWIW it could be outside of any class. What's the point of moving it from one class to another?
check_write_error is only relevant to MYSQL_BIN_LOG (in the way it's written). It ensures that a write error is set in thd (so it's closer to ensure_write_error). I have returned the declaration to where it was, maybe I should even remove the change from the initial refactoring commit. Also i think it's very much binlog-specific, and it's creepily designed, so want it to be kept closer to its specifics.
@@ -6544,7 +6562,8 @@ Event_log::flush_and_set_pending_rows_event(THD *thd, Rows_log_event* event,
if (writer.write(pending)) { set_write_error(thd, is_transactional); - if (check_write_error(thd) && cache_data && + if (dynamic_cast<MYSQL_BIN_LOG*>(this) &&
why?
+ MYSQL_BIN_LOG::check_write_error(thd) && cache_data &&
stmt_has_updated_non_trans_table(thd)) cache_data->set_incident();
Yeah, probably it's better to make a virtual function... but the reason was that incident events is also a binlog-specific thing. And that check_write_error is moved to MYSQL_BIN_LOG as static. I'll improve it.
{ DBUG_ENTER("MYSQL_BIN_LOG::check_write_error");
@@ -3806,9 +3815,9 @@ bool MYSQL_BIN_LOG::open_index_file(const char *index_file_name_arg, }
-bool Event_log::open(enum cache_type io_cache_type_arg) +bool Event_log::open(enum cache_type io_cache_type_arg, size_t buffer_size) { - bool error= init_io_cache(&log_file, -1, LOG_BIN_IO_SIZE, io_cache_type_arg, + bool error= init_io_cache(&log_file, -1, buffer_size, io_cache_type_arg,
why? it's only used by online alter, where would you need a different buffer size? only in your DBUG_EXECUTE_IF?
Yes, only there.
0, 0, MYF(MY_WME | MY_NABP |
MY_WAIT_IF_FULL));
log_state= LOG_OPENED;
@@ -7721,8 +7749,11 @@ static int binlog_online_alter_end_trans(THD *thd, bool all, bool commit) { DBUG_ASSERT(cache.cache_log.type != READ_CACHE); mysql_mutex_lock(binlog->get_log_lock()); - error= binlog->write_cache(thd, &cache.cache_log); + error= binlog->write_cache_raw(thd, &cache.cache_log); mysql_mutex_unlock(binlog->get_log_lock()); + + if (unlikely(error)) + binlog->set_write_error(my_errno);
shouldn't binlog->write_cache() call set_write_error() internally? why would the caller have to do it?
Wouldn't you bother if I'll move it to mf_iocache as smth like my_b_copy_cache?
diff --git a/sql/log.h b/sql/log.h index 83e323cb6fb..39512b01991 100644 --- a/sql/log.h +++ b/sql/log.h @@ -444,17 +445,20 @@ class Event_log: public MYSQL_LOG */ class Cache_flip_event_log: public Event_log { IO_CACHE alt_buf; +IF_DBUG(public:,)
yuck I don't see you using online_write_error or ref_count in DBUG macros
IO_CACHE *current, *alt;
you mean, you want "private" set back here? Sure, I don't access neither of the following outside.
std::atomic<uint> ref_count; + std::atomic<int> online_write_error; public:
Cache_flip_event_log() : Event_log(), alt_buf{}, - current(&log_file), alt(&alt_buf), ref_count(1) {} - bool open(enum cache_type io_cache_type_arg) + current(&log_file), alt(&alt_buf), ref_count(1), + online_write_error(0) {} + bool open(size_t buffer_size)
looks like Cache_flip_event_log::open() doesn't need any arguments at all
I don't want to put DBUG_EXECUTE_IF inside, because I'll have to either #include it in log.h, or move Cache_flip_event_log::open in some .cc.
diff --git a/sql/log_event.cc b/sql/log_event.cc index 10180a95d99..0c4c284a40f 100644 --- a/sql/log_event.cc +++ b/sql/log_event.cc @@ -761,16 +761,21 @@ Log_event::Log_event(const uchar *buf,
int Log_event::read_log_event(IO_CACHE* file, String* packet, const Format_description_log_event *fdle, - enum enum_binlog_checksum_alg checksum_alg_arg) + enum enum_binlog_checksum_alg checksum_alg_arg, + size_t max_allowed_packet) { ulong data_len; char buf[LOG_EVENT_MINIMAL_HEADER_LEN]; uchar ev_offset= packet->length(); #if !defined(MYSQL_CLIENT) - THD *thd=current_thd; - ulong max_allowed_packet= thd ? thd->slave_thread ? slave_max_allowed_packet - : thd->variables.max_allowed_packet - : ~(uint)0; + if (max_allowed_packet == 0)
may be better pass THD as an argument? or even correct value of max_allowed_packet. As
ulong max_allowed_packet(THD *thd) { return thd ? ... : ~(uint)0; }
Makes sense. That'll also remove the above ifdef, which I was unsure of, whether to remove it, or not.
on the second thought, do you need to ignore max_allowed_packet here for online ALTER? Is there no event size limit when the event is written into binlog?
I haven't seen such. If it's there, better to ignore it as well. -- Yours truly, Nikita Malyavin