Erratically behaving query needs optimization

Hello,
I'm running "PostgreSQL 11.2, compiled by Visual C++ build 1914,
64-bit" and I have a query that runs several times per user action
(9-10 times).
The query takes a long time to execute, specially at first, due to
cold caches I think, but the performance varies greatly during a run
of the application (while applying the said action by the user several
times).

My tables are only getting bigger with time, not much DELETEs and even
less UPDATEs as far as I can tell.

Problematic query:

EXPLAIN (ANALYZE,BUFFERS)
SELECT DISTINCT ON (results.attribute_id) results.timestamp,
results.data FROM results
    JOIN scheduler_operation_executions ON
scheduler_operation_executions.id = results.operation_execution_id
    JOIN scheduler_task_executions ON scheduler_task_executions.id =
scheduler_operation_executions.task_execution_id
WHERE scheduler_task_executions.device_id = 97
    AND results.data <> '<NullData/>'
    AND results.data IS NOT NULL
    AND results.object_id = 1955
    AND results.attribute_id IN (4, 5) -- possibly a longer list here
    AND results.data_access_result = 'SUCCESS'
ORDER BY results.attribute_id, results.timestamp DESC
LIMIT 2 -- limit by the length of the attributes list

In words: I want the latest (ORDER BY results.timestamp DESC) results
of a device (scheduler_task_executions.device_id = 97 - hence the
joins results -> scheduler_operation_executions ->
scheduler_task_executions)
for a given object and attributes with some additional constraints on
the data column. But I only want the latest attributes for which we
have results, hence the DISTINCT ON (results.attribute_id) and LIMIT.

First run: https://explain.depesz.com/s/qh4C
Limit  (cost=157282.39..157290.29 rows=2 width=54) (actual
time=44068.166..44086.970 rows=2 loops=1)
  Buffers: shared hit=215928 read=85139
  ->  Unique  (cost=157282.39..157298.20 rows=4 width=54) (actual
time=44068.164..44069.301 rows=2 loops=1)
        Buffers: shared hit=215928 read=85139
        ->  Sort  (cost=157282.39..157290.29 rows=3162 width=54)
(actual time=44068.161..44068.464 rows=2052 loops=1)
              Sort Key: results.attribute_id, results."timestamp" DESC
              Sort Method: quicksort  Memory: 641kB
              Buffers: shared hit=215928 read=85139
              ->  Gather  (cost=62853.04..157098.57 rows=3162
width=54) (actual time=23518.745..44076.385 rows=4102 loops=1)
                    Workers Planned: 2
                    Workers Launched: 2
                    Buffers: shared hit=215928 read=85139
                    ->  Nested Loop  (cost=61853.04..155782.37
rows=1318 width=54) (actual time=23290.514..43832.223 rows=1367
loops=3)
                          Buffers: shared hit=215928 read=85139
                          ->  Parallel Hash Join
(cost=61852.61..143316.27 rows=24085 width=4) (actual
time=23271.275..40018.451 rows=19756 loops=3)
                                Hash Cond:
(scheduler_operation_executions.task_execution_id =
scheduler_task_executions.id)
                                Buffers: shared hit=6057 read=85139
                                ->  Parallel Seq Scan on
scheduler_operation_executions  (cost=0.00..74945.82 rows=2482982
width=8) (actual time=7.575..15694.435 rows=1986887 loops=3)
                                      Buffers: shared hit=2996 read=47120
                                ->  Parallel Hash
(cost=61652.25..61652.25 rows=16029 width=4) (actual
time=23253.337..23253.337 rows=13558 loops=3)
                                      Buckets: 65536  Batches: 1
Memory Usage: 2144kB
                                      Buffers: shared hit=2977 read=38019
                                      ->  Parallel Seq Scan on
scheduler_task_executions  (cost=0.00..61652.25 rows=16029 width=4)
(actual time=25.939..23222.174 rows=13558 loops=3)
                                            Filter: (device_id = 97)
                                            Rows Removed by Filter: 1308337
                                            Buffers: shared hit=2977 read=38019
                          ->  Index Scan using
index_operation_execution_id_asc on results  (cost=0.43..0.51 rows=1
width=58) (actual time=0.191..0.191 rows=0 loops=59269)
                                Index Cond: (operation_execution_id =
scheduler_operation_executions.id)
                                Filter: ((data IS NOT NULL) AND (data
<> '<NullData/>'::text) AND (attribute_id = ANY ('{4,5}'::integer[]))
AND (object_id = 1955) AND (data_access_result = 'SUCCESS'::text))
                                Rows Removed by Filter: 0
                                Buffers: shared hit=209871
Planning Time: 29.295 ms
Execution Time: 44087.365 ms


Second run: https://explain.depesz.com/s/uy9f
Limit  (cost=157282.39..157290.29 rows=2 width=54) (actual
time=789.363..810.440 rows=2 loops=1)
  Buffers: shared hit=216312 read=84755
  ->  Unique  (cost=157282.39..157298.20 rows=4 width=54) (actual
time=789.361..789.535 rows=2 loops=1)
        Buffers: shared hit=216312 read=84755
        ->  Sort  (cost=157282.39..157290.29 rows=3162 width=54)
(actual time=789.361..789.418 rows=2052 loops=1)
              Sort Key: results.attribute_id, results."timestamp" DESC
              Sort Method: quicksort  Memory: 641kB
              Buffers: shared hit=216312 read=84755
              ->  Gather  (cost=62853.04..157098.57 rows=3162
width=54) (actual time=290.356..808.454 rows=4102 loops=1)
                    Workers Planned: 2
                    Workers Launched: 2
                    Buffers: shared hit=216312 read=84755
                    ->  Nested Loop  (cost=61853.04..155782.37
rows=1318 width=54) (actual time=238.313..735.472 rows=1367 loops=3)
                          Buffers: shared hit=216312 read=84755
                          ->  Parallel Hash Join
(cost=61852.61..143316.27 rows=24085 width=4) (actual
time=237.966..677.975 rows=19756 loops=3)
                                Hash Cond:
(scheduler_operation_executions.task_execution_id =
scheduler_task_executions.id)
                                Buffers: shared hit=6441 read=84755
                                ->  Parallel Seq Scan on
scheduler_operation_executions  (cost=0.00..74945.82 rows=2482982
width=8) (actual time=0.117..234.279 rows=1986887 loops=3)
                                      Buffers: shared hit=3188 read=46928
                                ->  Parallel Hash
(cost=61652.25..61652.25 rows=16029 width=4) (actual
time=236.631..236.631 rows=13558 loops=3)
                                      Buckets: 65536  Batches: 1
Memory Usage: 2144kB
                                      Buffers: shared hit=3169 read=37827
                                      ->  Parallel Seq Scan on
scheduler_task_executions  (cost=0.00..61652.25 rows=16029 width=4)
(actual time=0.132..232.758 rows=13558 loops=3)
                                            Filter: (device_id = 97)
                                            Rows Removed by Filter: 1308337
                                            Buffers: shared hit=3169 read=37827
                          ->  Index Scan using
index_operation_execution_id_asc on results  (cost=0.43..0.51 rows=1
width=58) (actual time=0.003..0.003 rows=0 loops=59269)
                                Index Cond: (operation_execution_id =
scheduler_operation_executions.id)
                                Filter: ((data IS NOT NULL) AND (data
<> '<NullData/>'::text) AND (attribute_id = ANY ('{4,5}'::integer[]))
AND (object_id = 1955) AND (data_access_result = 'SUCCESS'::text))
                                Rows Removed by Filter: 0
                                Buffers: shared hit=209871
Planning Time: 1.787 ms
Execution Time: 810.634 ms

You can see that the second run takes less than one second to run...
which is 43 seconds better than the first try, just by re-running the
query.
Other runs take maybe 1s, 3s, still a long time.

How can I improve it to be consistently fast (is it possible to get to
several milliseconds?)?
What I don't really understand is why the nested loop has 3 loops
(three joined tables)?
And why does the first index scan indicate ~60k loops? And does it
really work? It doesn't seem to filter out any rows.

Should I add an index only on (attribute_id, object_id)? And maybe
data_access_result?
Does it make sens to add it on a text column (results.data)?

My tables:

https://gist.githubusercontent.com/paulbarbu/0d36271d710349d8fb6102d9a466bb54/raw/7a6946ba7c2adec5b87ed90f343f1aff37432d21/gistfile1.txt

As you can see from the gist the foreign keys are indexed. Other
indices were added to speed up other queries.
Other relevant information (my tables have 3+ millions of rows, not
very big I think?), additional info with regards to size also included
below.
This query has poor performance on two PCs (both running off of HDDs)
so I think it has more to do with my indices and query than Postgres
config & hardware, will post those if necessary.


Size info:
SELECT relname, relpages, reltuples, relallvisible, relkind, relnatts,
relhassubclass, reloptions, pg_table_size(oid) FROM pg_class WHERE
relname IN ('results', 'scheduler_operation_executions',
'scheduler_task_executions');
-[ RECORD 1 ]--+-------------------------------
relname        | results
relpages       | 65922
reltuples      | 3.17104e+06
relallvisible  | 65922
relkind        | r
relnatts       | 9
relhassubclass | f
reloptions     |
pg_table_size  | 588791808
-[ RECORD 2 ]--+-------------------------------
relname        | scheduler_operation_executions
relpages       | 50116
reltuples      | 5.95916e+06
relallvisible  | 50116
relkind        | r
relnatts       | 8
relhassubclass | f
reloptions     |
pg_table_size  | 410697728
-[ RECORD 3 ]--+-------------------------------
relname        | scheduler_task_executions
relpages       | 40996
reltuples      | 3.966e+06
relallvisible  | 40996
relkind        | r
relnatts       | 12
relhassubclass | f
reloptions     |
pg_table_size  | 335970304

Thanks for your time!

--
Barbu Paul - Gheorghe