|
| 1 | +#coding:utf-8 |
| 2 | + |
| 3 | +""" |
| 4 | +ID: issue-7687 |
| 5 | +ISSUE: https://github.com/FirebirdSQL/firebird/issues/7687 |
| 6 | +TITLE: Add LEVEL column to PLG$PROF_RECORD_SOURCES and PLG$PROF_RECORD_SOURCE_STATS_VIEW |
| 7 | +DESCRIPTION: |
| 8 | + Test creates master-detail tables and view 'v_test' for them. |
| 9 | + Then we create auxiliary view 'vp_rec_stats' to see profiler data. |
| 10 | + Running query to 'v_test' will add data to profiler snapshot tables so that we must have ability |
| 11 | + to reconstruct the whole EXPLAINED PLAN of that query, i.e. with top-level and all subsequent access paths. |
| 12 | + This reconstruction is performed by recursive query which first gets records with 'parent_record_source_id is null' |
| 13 | + (and all these records always have LEVEL = 0), and then goes down according to the principle: |
| 14 | + "give be lines with level that be bigger but CLOSEST to current one". This is achieved by applying dense_rank() |
| 15 | + function, see 'ranked_level' column usage and notes below. |
| 16 | +NOTES: |
| 17 | + [25.08.2023] pzotov |
| 18 | + 1. We can NOT use "next level must be current one plus 1" in recursive part of query that reconstructs explained plan! |
| 19 | + This is because some access paths can be 'splitted' onto one or several 'sub-paths' by LF character, thus they are looked |
| 20 | + to be somewhat as 'multi-line' parts. Every LF in such case causes LEVEL to be "sinlently" increased by 1, so lines |
| 21 | + AFTER this "complex" access path will have LEVEL that differs from source for MORE than 1. |
| 22 | +
|
| 23 | + Here is example of such data (there is NO record with LEVEL = 3): |
| 24 | + STTM_ID LEVEL REC_ID PAR_ID ACC_PATH SQL_TEXT |
| 25 | + ======= ======= ======= ======= ================================================== ================================================== |
| 26 | + 5760 0 1 <null> Select Expression select first 1 r.* from rdb$database r order by 1 |
| 27 | + 5760 1 2 1 -> First N Records select first 1 r.* from rdb$database r order by 1 |
| 28 | + 5760 2 3 2 -> Refetch |
| 29 | + -> Sort (record length: 36, key select first 1 r.* from rdb$database r order by 1 |
| 30 | + 5760 4 4 3 -> Table "RDB$DATABASE" as "R" Full Scan select first 1 r.* from rdb$database r order by 1 |
| 31 | +
|
| 32 | + This was implemented intentionally with purpose to make reconstruction of indents easier. |
| 33 | + Because of that, view that queries PLG$PROF_ tables must use "dense_rank()over(order by t.level)". |
| 34 | + 2. We have to make additional (preliminary) call to rdb$profiler.start_session() / .finish_session() BEFORE test in order to create view |
| 35 | + that is based on profiler snapshot tables. Otherwise error raises: |
| 36 | + SQLSTATE = 42S02 / ... / -Table unknown / -PLG$PROF_RECORD_SOURCE_STATS_VIEW |
| 37 | + See: doc/sql.extensions/README.profiler.md |
| 38 | + "Snapshot tables (as well views and sequence) are automatically created in the first usage of the profiler" |
| 39 | + 3. Currently there is problem with distinguish 'blocks' from profiler data that corresponds to sub-queries: there is no column that applies |
| 40 | + to the whole block and lines that belong (by hierarchy) to subquery can be displayed either before or after lines that relate to 'main' |
| 41 | + part of query. Although it seems that subquery is mostly displayed *after* main part. |
| 42 | + Because of that, it was decided for now to leave expected output as it is produced by current FB versions. |
| 43 | + If future runs will show that subquery part "jumps" randomly (from "top" to "bottom") than only presense of top-level access paths |
| 44 | + will be checked (without requirement to their concrete order). |
| 45 | + 4. In order to see indents properly (including case if they will change in size) we have to prefix each line ('#' character is used). |
| 46 | +
|
| 47 | + Checked on 5.0.0.1169. |
| 48 | +""" |
| 49 | + |
| 50 | +import os |
| 51 | +import pytest |
| 52 | +from firebird.qa import * |
| 53 | + |
| 54 | +db = db_factory() |
| 55 | +act = python_act('db') #, substitutions=[('[ \t]+', ' ')]) |
| 56 | + |
| 57 | +@pytest.mark.version('>=5.0') |
| 58 | +def test_1(act: Action, capsys): |
| 59 | + |
| 60 | + test_sql = f""" |
| 61 | + create table tmain(id int primary key using index tmain_pk, x int); |
| 62 | + create table tdetl(id int primary key using index tdetl_pk, pid int references tmain using index tdetl_fk, y int, z int); |
| 63 | + insert into tmain(id,x) select row_number()over(), -100 + rand()*200 from rdb$types rows 100; |
| 64 | + insert into tdetl(id, pid, y,z) select row_number()over(), 1+rand()*99, rand()*1000, rand()*1000 from rdb$types; |
| 65 | + commit; |
| 66 | + create index tmain_x on tmain(x); |
| 67 | + create index tdetl_y on tdetl(y); |
| 68 | + create index tdetl_z on tdetl(z); |
| 69 | +
|
| 70 | + set statistics index tdetl_fk; |
| 71 | + commit; |
| 72 | + |
| 73 | + out nul; |
| 74 | + -- This is needed in order to create view based on snapshot 'plg$prof_*' tables: |
| 75 | + select rdb$profiler.start_session('profile session 0') from rdb$database; |
| 76 | + out; |
| 77 | +
|
| 78 | + create view vp_rec_stats as |
| 79 | + select |
| 80 | + cast(t.statement_id as smallint) as sttm_id |
| 81 | + ,cast(t.level as smallint) as level |
| 82 | + ,cast(t.record_source_id as smallint) as rec_id |
| 83 | + ,cast(t.parent_record_source_id as smallint) as par_id |
| 84 | + ,substring('#' || lpad('', 4*t.level,' ') || replace( replace(t.access_path, ascii_char(13), ''), ascii_char(10), ascii_char(10) || '#' || lpad('', 4*t.level,' ') ) from 1 for 320) as acc_path |
| 85 | + ,substring( cast(t.sql_text as varchar(255)) from 1 for 50 ) as sql_text |
| 86 | + ,dense_rank()over(order by t.level) as ranked_level |
| 87 | + from plg$prof_record_source_stats_view t |
| 88 | + join plg$prof_sessions s |
| 89 | + on s.profile_id = t.profile_id and |
| 90 | + s.description = 'profile session 1' |
| 91 | + ; |
| 92 | + execute procedure rdb$profiler.finish_session(true); |
| 93 | + commit; |
| 94 | +
|
| 95 | +
|
| 96 | + create view v_test as |
| 97 | + select m4.id, d4.y, d4.z |
| 98 | + from tmain m4 |
| 99 | + cross join lateral ( |
| 100 | + select y, z |
| 101 | + from tdetl dx |
| 102 | + where |
| 103 | + dx.pid = m4.id |
| 104 | + and m4.x between dx.y and dx.z |
| 105 | + ) d4 |
| 106 | + where exists(select count(*) from tdetl dy group by dy.pid having count(*) > 2); |
| 107 | +
|
| 108 | + out nul; |
| 109 | + select rdb$profiler.start_session('profile session 1') from rdb$database; |
| 110 | + -- Test query for which we want to see data in the profiler tables: |
| 111 | + -- ########## |
| 112 | + select * from v_test; |
| 113 | + out; |
| 114 | + execute procedure rdb$profiler.finish_session(true); |
| 115 | + commit; |
| 116 | +
|
| 117 | + set transaction read committed; |
| 118 | + set count on; |
| 119 | + set list off; |
| 120 | + set heading off; |
| 121 | +
|
| 122 | + -- ############################################## |
| 123 | + -- Output data from profiler. |
| 124 | + -- NB-1: each line is prefixed with '#' character in order to see indents. |
| 125 | + -- NB-2: 'level' column can not be used for joining in recursive part because of 'gaps'! |
| 126 | + -- Instead, 'ranked_level' is used fo this: |
| 127 | + -- ############################################## |
| 128 | + with recursive |
| 129 | + r as ( |
| 130 | + select sttm_id, level, t.ranked_level, rec_id, par_id, acc_path --, sql_text |
| 131 | + from vp_rec_stats t |
| 132 | + where |
| 133 | + t.par_id is null and t.sql_text not containing 'rdb$profiler.start_session' |
| 134 | + |
| 135 | + UNION ALL |
| 136 | +
|
| 137 | + select t.sttm_id, t.level, t.ranked_level, t.rec_id, t.par_id, t.acc_path --, t.sql_text |
| 138 | + from vp_rec_stats t |
| 139 | + join r on t.sttm_id = r.sttm_id and t.ranked_level = r.ranked_level + 1 and r.rec_id = t.par_id |
| 140 | + ) |
| 141 | + select acc_path from r; |
| 142 | + """ |
| 143 | + |
| 144 | + act.expected_stdout = f""" |
| 145 | + #Select Expression |
| 146 | + # -> Filter (preliminary) |
| 147 | + # -> Nested Loop Join (inner) |
| 148 | + # -> Table "TMAIN" as "V_TEST M4" Full Scan |
| 149 | + # -> Filter |
| 150 | + # -> Table "TDETL" as "V_TEST D4 DX" Access By ID |
| 151 | + # -> Bitmap And |
| 152 | + # -> Bitmap |
| 153 | + # -> Index "TDETL_FK" Range Scan (full match) |
| 154 | + # -> Bitmap |
| 155 | + # -> Index "TDETL_Y" Range Scan (upper bound: 1/1) |
| 156 | + #Sub-query (invariant) |
| 157 | + # -> Filter |
| 158 | + # -> Aggregate |
| 159 | + # -> Table "TDETL" as "V_TEST DY" Access By ID |
| 160 | + # -> Index "TDETL_FK" Full Scan |
| 161 | +
|
| 162 | + Records affected: 10 |
| 163 | + """ |
| 164 | + act.isql(input = test_sql, combine_output = True) |
| 165 | + assert act.clean_stdout == act.clean_expected_stdout |
| 166 | + act.reset() |
0 commit comments