pg_stat_monitor is a bit longer; therefore, it requires some code cleanup.
Therefore I decided to turn these tasks into multiple commits and PR to avoid
various changes in one PR. This will ease the review and Q/A process.
In this commit, I have done these tasks.
1 - Remove all benchmarking and debugging code.
PG-456: Running pgindent to make source indentation/spacing PostgreSQLCompatible.
PostgreSQL uses pgindent from time to time to make source code PostgreSQL
style guide compatible, it is a very long time since we have not done that.
Commit fixes a lot of indentation and spacing issues.
Co-authored-by: Hamid Akhtar <hamid.akhtar@gmail.com>
The buckets are now created with the start time a modulus of the bucket time size.
So if we have a 10 second bucket, the start times would reflect that:
- Bucket1: 00:00:00
- Bucket2: 00:00:10
- Bucket3: 00:00:20
...
Previously, the start time of the bucket was aligned with the first query that
arrives in that bucket. However, now the behaviour is changed. So, even if the
first query for bucket 2 arrives at 00:00:13, the start time would still be set
to 00:00:10.
This change now makes the bucketing separated out by fixed time windows so that
external applications can easily consume that data and chart it.
Also, as part of this change, locking of pgss is updated now and extended
to last the bucket related changes.
There was no maximum limit set for the number of maximum histograms
bucket, which can lead to a crash in case of higher value.
PG-382: Adjust the maximum value for histogram buckets.
Fix the regression issue related to GUC and set the maximum
buckets value correctly.
PG-382: Adjust the maximum value for histogram buckets.
Fix the TAP test cases.
If pgsm_overflow_target is ON (default, 1) and the query buffer
overflows, we now dump the buffer and keep track of how many times
pg_stat_monitor changed bucket since that.
If an overflow happen again before pg_stat_monitor cycle through
pgsm_max_buckets buckets (default 10), then we don't dump the buffer
again, but instead report an error, this ensures that only one dump file
of size pgsm_query_shared_buffer will be in disk at any time, avoiding
slowing down queries to the pg_stat_monitor view.
As soon as pg_stat_monitor cycles through all buckets, we remove the
dump file and reset the counter (pgss->n_bucket_cycles).
pgss_ExecutorEnd: Avoid unnecessary memset(plan_info, 0, ...).
We only use this object if the condition below is true, in which case we
already initialize all the fields in the object, also we now store the
plan string length (plan_info.plan_len) to avoid calling strlen on it
again later:
if (queryDesc->operation == CMD_SELECT && PGSM_QUERY_PLAN) {
... here we initialize plan_info
If the condition is false, then we pass a NULL PlanInfo* to the
pgss_store to avoid more unnecessary processing.
pgss_planner_hook: Similar, avoid memset(plan_info, 0, ...) this object
is not used here, so we pass NULL to pgss_store.
pg_get_application_name: Remove call to strlen, snprintf already give us
the calculated string length, so we just return it.
pg_get_client_addr: Cache localhost, avoid calling
ntohl(inet_addr("127.0.0.1")) all the time.
pgss_update_entry: Make use of PlanInfo->plan_len, avoiding a call to
strlen again.
intarray_get_datum: Init the string by setting the first byte to '\0'.
The memory area reserved for query text (pgsm_query_shared_buffer) was
divided evenly for each bucket, this allowed to have the same query,
e.g. "SELECT 1", duplicated in different buckets, thus wasting space.
This commit fix the query text duplication by adding a new hash table
whose only purpose is to verify if a given query is already added to the
buffer (by using the queryID).
This allows different buckets that share the same query to point to a
unique entry in the query buffer (pgss_qbuf).
When pg_stat_monitor moves to a new bucket id, by avoiding adding a
query that already exists in the buffer it can also save some CPU time.
This commit fix some issues when the query buffer overflows and
pg_stat_monitor attempts to dump its contents to a file.
The dump process is now as follows:
1. The dump will always be a full dump of the current query buffer,
meaning pg_stat_monitor will dump MAX_QUERY_BUFFER_BUCKET bytes to
the dump file.
2. When scanning the dump file, read chunks of size
MAX_QUERY_BUFFER_BUCKET, then look for the query ID using that chunk
and the query position metadata, this allows pg_stat_monitor to avoid
scanning the whole chunk when looking for a query ID.
The code in charge to read from/write to the dump file now takes into
account that read() and write() may return less bytes than what it was
asked for, the code now ensures that we actually read or write the
amount of bytes required (MAX_QUERY_BUFFER_BUCKET), also it handles
rare but posssible interrupts when doing those operations.
The GUC variable pgsm_overflow_target was pointing to the wrong index
(12, pgsm_track_planning) in the "GucVariable conf[MAX_SETTINGS]" array.
Adjusted it to the right index, 11.
Whenever a new query is added to a query buffer, record the position
in which the query was inserted, we can then use this information to
locate the query in a faster way later on when required.
This allowed to simplify the logic in hash_entry_dealloc(), after
creating the list of pending queries, as the list is scanned we can copy
the query from the previous query buffer to the new one by using the
query position (query_pos), this avoids scanning the whole query buffer
when looking up for the queryid.
Also, when moving a query to a new buffer (copy_query), we update
the query_pos for the hash table entry, so it points to the right place
in the new query buffer.
read_query() function was updated to allow query position to be passed
as argument, if pos != 0 use it to locate the query directly, otherwise
fallback to the previous mode of scanning the whole buffer.
SaveQueryText() was updated to pass a reference to the query position as
argument, this value is updated after the function finishes with the
position that the query was stored into the buffer.
There was no necessity for using a separate hash table to keep track
of queries as all the necessary information is already available in
the pgss_hash table.
The code that moves pending queries' text in hash_query_entry_dealloc
was merged into hash_entry_dealloc.
Couple functions were not necessary anymore, thus were removed:
- hash_create_query_entry
- pgss_store_query_info
- pgss_get_entry (this logic was added directly into pgss_store).
We simplified the logic in pgss_store, it basically works as follows:
1. Create a key (bucketid, queryid, etc...) for the query event.
2. Lookup the key in pgss_hash, if no entry exists for the key, create a
new one, save the query text into the current query buffer.
3. If jstate == NULL, then update stats counters for the entry.
Added code to move all pending queries text from an expired bucket's query
buffer to the next, active query buffer.
The previous implementation was not very efficient, it worked like this,
as soon as a query is processed and a bucket expires:
1. Allocate memory to save the contents of the next query buffer.
2. Clear the next query buffer.
3. Iterate over pgss_query_hash, then, for each entry:
- If the entry's bucket id is equal to the next bucket then:
-- If the query for this entry has finished or ended in error, then
remove it from the hash table.
-- Else, if the query is not yet finished, copy the query from the
backup query buffer to the new query buffer.
Now, this copy was really expensive, because it was implemented
using read_query() / SaveQueryText(), and read_query() scans the
whole query buffer looking for some query ID, since we do this
extra lookup loop for each pending query we end up with a O(n^2)
algorithm.
4. Release the backup query buffer.
Since now we always move pending queries from an expired bucket to the
next one, there is no need to scan the next query buffer for pending
queries (the pending queries are always in the current bucket, and when
it expires we move them to the next one).
Taking that into consideration, the new implementation works as follows,
whenever a bucket expires:
1. Clear the next query buffer (all entries).
2. Define an array to store pending query ids from the expired bucket,
we use this array later on in the algorithm.
3. Iterate over pgss_query_hash, then, for each entry:
- If the entry's bucket id is equal to the next bucket then:
-- If the query for this entry has finished or ended in error, then
remove it from the hash table. This is equal to the previous
implementation.
- Else, if the entry's bucket id is equal to the just expired bucket
id (old bucket id) and the query state is pending (not yet finished),
then add this query ID to the array of pending query IDs.
Note: We define the array to hold up to 128 pending entries, if there
are more entries than this we try to allocate memory in the heap to
store them, then, if the allocation fails we manually copy every
pending query past the 128th to the next query buffer, using the
previous algorithm (read_query() / SaveQueryText), this should be a
very rare situation.
4. Finally, if there are pending queries, copy them from the previous
query buffer to the next one using copy_queries.
Now, copy_queries() is better than looping through each query entry and
calling read_query() / SaveQueryText() to copy each of them to the new
buffer, as explained, read_query() scans the whole query buffer for
every call.
copy_queries(), instead, scans the query buffer only once, and for every
element it checks if the current query id is in the list of queries to
be copied, this is an array of uint64 that is small enough to fit in L1
cache.
Another important fix in this commit is the addition of the line 1548 in
pg_stat_monitor.c / pgss_store():
query_entry->state = kind;
Before the addition of this line, all entries in the pgss_query_hash
hash table were not having their status updated when the query entered
execution / finished or ended in error, effectively leaving all entries
as pending, thus whenever a bucket expired all entries were being copied
from the expired bucket to the next one.
The if condition bellow in geta_next_wbucket() was subject to a race
condition:
if ((current_usec - pgss->prev_bucket_usec) > (PGSM_BUCKET_TIME * 1000 *
1000))
Two or more threads/processes could easily evaluate this condition to
true, thus executing more than once the block that would calculate a
new bucket id, clear/move old entries in the pgss_query_hash and
pgss_hash hash tables.
To avoid this problem, we define prev_bucket_usec and current_wbucket
variables as atomic and execute a loop to check if another thread has
updated prev_bucket_usec before the current one.
A problem during bucket management was allowing some queries to be
duplicated, old entries would sit around without having their statistics
updated.
The problem would arise during the following chain of events:
1. A query goes through pgss_post_parse_analyze, in this stage (PGSS_PARSE)
we only save the query into the query buffer and create an entry in the
query hash table.
2. The query then goes through pgss_ExecutorStart (PGSS_EXEC), in this stage
we create an entry for query statistic counters with default values,
all time stats equal zero, etc.
3. The query then goes through pgss_ExecutorEnd (PGSS_FINISH), in this stage
we update the query statistis, no. calls, total time taken, min_time, etc.
The problem is that between steps 2 and 3, the current bucket ID timer may
have been expired.
For example, during steps 1 and 2 the query may have been stored in
bucket ID 1, but when the query is finished (pgss_ExecutorEnd) the
current bucket ID may have been updated to 2.
This is leaving an entry for the query in bucket ID 1 with state ACTIVE,
with time statistics not updated yet.
This is also creating an entry for the query in the bucket ID 2, with all
statistics (time and others) being updated for this entry.
To solve this problem, during transition to a new bucket id, we scan all
pending queries in the previous bucket id and move them to the new
bucket id.
This way finished queries will always be associated with the bucket id
that was active at the time they've finished.
These variables can't be in shared state, as the following problem was
taking place:
1. Process1 call pgss_ExecutorCheckPerms(), acquire lock, update
relations and num_relations, release lock.
2. Process 2 call pgss_ExecutorCheckPerms(), acquire lock, update
num_relations = 0;
3. Process 1 read num_relations = 0 in pgss_update_entry, this value is
wrong as it was updated by Process 2.
Even if we acquire the lock in pgss_update_entry to read num_relations
and relations variable, Process 1 may end up acquiring the lock after
Process 2 has ovewritten the variable values, leading to Process 1
reading of wrong data.
By defining relations and num_relations to be static and global in
pg_stat_monitor.c we take advantage that each individual PostgreSQL
backend will have its own copy of this data, which allows us to remove
the locking in pgss_ExecutorCheckPerms to update these variables,
improving pg_stat_monitor overall performance.
Added a new view 'pg_stat_monitor_hook_stats' that provide execution
time statistics for all hooks installed by the module, following is a
description of the fields:
- hook: The hook function name.
- min_time: The fastest execution time recorded for the given hook.
- max_time: The slowest execution time recorded for the given hook.
- total_time: Total execution time taken by all calls to the hook.
- avg_time: Average execution time of a call to the hook.
- ncalls: Total number of calls to the hook.
- load_comparison: A percentual of time taken by an individual hook
compared to every other hook.
To enable benchmark, code must be compiled with -DBENCHMARK flag, this
will make the hook functions to be replaced by a function with the same
name plus a '_benchmark' suffix, e.g. hook_function_benchmark.
The hook_function_benchmark will call the original function and
calculate the amount of time it took to execute, than it will update
statistics for that hook.
Add application name to the key used to identify queries in the hash
table, this allows different applications to have separate entries in
pg_stat_monitor view if they issued the same query.
Ibrar Ahmed
Hamid Akhtar
Ibrar Ahmed
Diego Fronza
Andrew Pogrebnoy