1. Input and output RowGroup's used in GROUP_CONCAT classes
are currently allocating a raw memory buffer of size equal
to the actual width of the string datatype. As an example,
for the following query:
SELECT col1, GROUP_CONCAT(col2) FROM t GROUP BY col1;
If col2 is a TEXT field with default width, the input
RowGroup containing the target rows to be concatenated will
assign 64kb of memory for every input row in the RowGroup.
This is wasteful as actual field values in real workloads
would be much smaller. We fix this by enabling the
RowGroup to use the StringStore when the RowGroup contains
long strings.
2. RowAggregation::initialize() allocates a memory buffer
for a NULL row. The size of this buffer is equal to the
row size for the output RowGroup. For the above scenario,
using the default group_concat_max_len (which is a server
variable that sets the maximum length of the GROUP_CONCAT string)
value of 1mb, the buffer size would be
(1mb + 64kb + some additional metadata). If the user sets
group_concat_max_len to a higher value, say 3gb, this buffer
size would be ~3gb. Now if the runtime initiates several
instances of RowAggregation, total memory consumption by
PrimProc could exceed the hardware memory limits causing the
OS OOM to kill the process. We fix this problem by again
enabling the StringStore for the NULL row allocation.
3. In the plugin code in buildAggregateColumn(), there is
an integer overflow when the server group_concat_max_len
variable (which is an uint32_t) is set to a value > INT32_MAX
(such as 3gb) and is assigned to
CalpontSystemCatalog::ColType::colWidth (which is an int32_t).
As a short term fix, we saturate the assigned value to colWidth
to INT32_MAX. Proper fix would be to upgrade
CalpontSystemCatalog::ColType::colWidth to an uint32_t.
Part 1:
As part of MCOL-3776 to address synchronization issue while accessing
the fTimeZone member of the Func class, mutex locks were added to the
accessor and mutator methods. However, this slows down processing
of TIMESTAMP columns in PrimProc significantly as all threads across
all concurrently running queries would serialize on the mutex. This
is because PrimProc only has a single global object for the functor
class (class derived from Func in utils/funcexp/functor.h) for a given
function name. To fix this problem:
(1) We remove the fTimeZone as a member of the Func derived classes
(hence removing the mutexes) and instead use the fOperationType
member of the FunctionColumn class to propagate the timezone values
down to the individual functor processing functions such as
FunctionColumn::getStrVal(), FunctionColumn::getIntVal(), etc.
(2) To achieve (1), a timezone member is added to the
execplan::CalpontSystemCatalog::ColType class.
Part 2:
Several functors in the Funcexp code call dataconvert::gmtSecToMySQLTime()
and dataconvert::mySQLTimeToGmtSec() functions for conversion between seconds
since unix epoch and broken-down representation. These functions in turn call
the C library function localtime_r() which currently has a known bug of holding
a global lock via a call to __tz_convert. This significantly reduces performance
in multi-threaded applications where multiple threads concurrently call
localtime_r(). More details on the bug:
https://sourceware.org/bugzilla/show_bug.cgi?id=16145
This bug in localtime_r() caused processing of the Functors in PrimProc to
slowdown significantly since a query execution causes Functors code to be
processed in a multi-threaded manner.
As a fix, we remove the calls to localtime_r() from gmtSecToMySQLTime()
and mySQLTimeToGmtSec() by performing the timezone-to-offset conversion
(done in dataconvert::timeZoneToOffset()) during the execution plan
creation in the plugin. Note that localtime_r() is only called when the
time_zone system variable is set to "SYSTEM".
This fix also required changing the timezone type from a std::string to
a long across the system.
* Introduce multigeneration aggregation
* Do not save unused part of RGDatas to disk
* Add IO error explanation (strerror)
* Reduce memory usage while aggregating
* introduce in-memory generations to better memory utilization
* Try to limit the qty of buckets at a low limit
* Refactor disk aggregation a bit
* pass calculated hash into RowAggregation
* try to keep some RGData with free space in memory
* do not dump more than half of rowgroups to disk if generations are
allowed, instead start a new generation
* for each thread shift the first processed bucket at each iteration,
so the generations start more evenly
* Unify temp data location
* Explicitly create temp subdirectories
whether disk aggregation/join are enabled or not
mcsconfig.h and my_config.h have the following
pre-processor definitions:
1. Conflicting definitions coming from the standard cmake definitions:
- PACKAGE
- PACKAGE_BUGREPORT
- PACKAGE_NAME
- PACKAGE_STRING
- PACKAGE_TARNAME
- PACKAGE_VERSION
- VERSION
2. Conflicting definitions of other kinds:
- HAVE_STRTOLL - this is a dirt in MariaDB headers.
Should be fixed in the server code. my_config.h erroneously
performs "#define HAVE_STRTOLL" instead of "#define HAVE_STRTOLL 1".
in some cases. The former is not CMake compatible style. The latter is.
3. Non-conflicting definitions:
Otherwise, mcsconfig.h and my_config.h should be mutually compatible,
because both are generated by cmake on the same host machine. So
they should have exactly equal definitions like "HAVE_XXX", "SIZEOF_XXX", etc.
Observations:
- It's OK to include both mcsconfig.h and my_config.h providing that we
suppress duplicate definition of the above conflicting types #1 and #2.
- There is no a need to suppress duplicate definitions mentioned in #3,
as they are compatible!
- my_sys.h and m_ctype.h must always follow a CMake configuation header,
either my_config.h or mcsconfig.h (or both).
They must never be included without any preceeding configuration header.
This change make sure that we resolve conflicts by:
- either disallowing inclusion of mcsconfig.h and my_config.h
at the same time
- or by hiding conflicting definitions #1 and #2
(with their later restoring).
- also, by making sure that my_sys.h and m_ctype.h always follow
a CMake configuration file.
Details:
- idb_mysql.h can now only be included only after my_config.h
An attempt to use idb_mysql.h with mcsconfig.h instead of
my_config.h is caught by the "#error" preprocessor directive.
- mariadb_my_sys.h can now be only included after mcsconfig.h.
An attempt to use mariadb_my_sys.h without mcscofig.h
(e.g. with my_config.h) is also caught by "#error".
- collation.h now can now be included in two ways.
It now has the following effective structure:
#if defined(PREFER_MY_CONFIG_H) && defined(MY_CONFIG_H)
// Remember current conflicting definitions on the preprocessor stack
// Undefine current conflicting definitions
#endif
#include "mcsconfig.h"
#include "m_ctype.h"
#if defined(PREFER_MY_CONFIG_H) && defined(MY_CONFIG_H)
# Restore conflicting definitions from the preprocessor stack
#endif
and can be included as follows:
a. using only mcsconfig.h as a configuration header:
// my_config.h must not be included so far
#include "collation.h"
b. using my_config.h as the first included configuration file:
#define PREFER_MY_CONFIG_H // Force conflict resolution
#include "my_config.h" // can be included directly or indirectly
...
#include "collation.h"
Other changes:
- Adding helper header files
utils/common/mcsconfig_conflicting_defs_remember.h
utils/common/mcsconfig_conflicting_defs_restore.h
utils/common/mcsconfig_conflicting_defs_undef.h
to perform conflict resolution easier.
- Removing `#include "collation.h"` from a number of files,
as it's automatically included from rowgroup.h.
- Removing redundant `#include "utils_utf8.h"`.
This change is not directly related to the problem being fixed,
but it's nice to remove redundant directives for both collation.h
and utils_utf8.h from all the files that do not really need them.
(this change could probably have gone as a separate commit)
- Changing my_init() to MY_INIT(argv[0]) in the MCS services sources.
After the fix of the complitation failure it appeared that ColumnStore
services compiled with the debug build crash due to recent changes in
safemalloc. The crash happened in strcmp() with `my_progname` as an argument
(where my_progname is a mysys global variable). This problem should
probably be fixed on the server side as well to avoid passing NULL.
But, the majority of MariaDB executable programs also use MY_INIT(argv[0])
rather than my_init(). So let's make MCS do like the other programs do.
After a UDAF result has been inserted in the output stream, the valOut object needs to be reset to empty in preparation for the next value. Failing to do so may cause what should be a NULL value to erroneously take the last value inserted.
Removed uint128 from joblist/lbidlist.*
Another toString() method for wide-decimal that is EMPTY/NULL aware
Unified decimal processing in WF functions
Fixed a potential issue in EqualCompData::operator() for
wide-decimal processing
Fixed some signedness warnings
For now it consists of only:
using int128_t = __int128;
using uint128_t = unsigned __int128;
All new privitive data types should go into this file in the future.
Binary NULL magic now consists of a series of BINARYEMPTYROW-s + BINARYNULL
in the end.
ByteStream now has hexbyte alias.
Added ColumnCommand::getEmptyRowValue to support 16 byte EMPTY values.
Replaced BINARYEMPTYROW and BINARYNULL values. We need to have
separate magic values for numeric and non-numeric binary types
b/c numeric cant tolerate losing 0 used for magics previously.
atoi128() now parses minus sign and produces negative values.
RowAggregation::isNull() now uses Row::isNull() for DECIMAL.
TupleAggregateStep class method and buildAggregateColumn() now properly set result data type.
doSum() now handles DECIMAL(38) in approprate manner.
Low-level null related methods for new binary-based datatypes now handles magic values for
binary-based DT.
