This is in a stored procedure..This if statement, then I do a little work. The #AsOfDate is a passed in variable of date datatype. The question I have is Why do I get better performance by removing the inner-most exists, but ONLY when the entire statement is in an IF EXISTS?
The two tables:
dbo.TXXX_InventoryDetail -- 1.3 billion records..stats up to date
dbo.TXXX_InventoryFull -- 9.8 million records..stats up to date
Statement:
if exists (select 1
from dbo.TXXX_InventoryDetail o
where exists (select 1
from dbo.TXXX_InventoryFull i
where i.C001_AsOfDate= o.C001_AsOfDate
and i.C001_ProductID=o.C001_ProductID
and i.C001_StoreNumber=o.C001_StoreNumber
and i.C001_AsOfDate=#AsOfDate
and (i.C001_LastModelDate!=o.C001_LastModelDate
or o.C001_InventoryQty!=o.C001_InventoryQty
or i.C001_OnOrderQty!=o.C001_OnOrderQty
or i.C001_TBOQty!=o.C001_TBOQty
or i.C001_ModelQty!=o.C001_ModelQty
or i.C001_TBOAdjustQty!=o.C001_TBOAdjustQty
or i.C001_ReturnQtyPending!=o.C001_ReturnQtyPending
or i.C001_ReturnQtyInProcess!=o.C001_ReturnQtyInProcess
or i.C001_ReturnQtyDueOut!=o.C001_ReturnQtyDueOut))
and o.C001_AsOfDate=#AsOfDate)
io output:
Table 'TXXX_InventoryFull'. Scan count 9240262, logical reads 29548864
Table 'T001_InventoryDetail'. Scan count 1, logical reads 17259
If I remove the second where exists and do a join:
if exists (select 1
from dbo.TXXX_InventoryDetail o,
dbo.TXXX_InventoryFull i
where i.C001_AsOfDate= o.C001_AsOfDate
and i.C001_ProductID=o.C001_ProductID
and i.C001_StoreNumber=o.C001_StoreNumber
and i.C001_AsOfDate=#AsOfDate
and (i.C001_LastModelDate!=o.C001_LastModelDate
or o.C001_InventoryQty!=o.C001_InventoryQty
or i.C001_OnOrderQty!=o.C001_OnOrderQty
or i.C001_TBOQty!=o.C001_TBOQty
or i.C001_ModelQty!=o.C001_ModelQty
or i.C001_TBOAdjustQty!=o.C001_TBOAdjustQty
or i.C001_ReturnQtyPending!=o.C001_ReturnQtyPending
or i.C001_ReturnQtyInProcess!=o.C001_ReturnQtyInProcess
or i.C001_ReturnQtyDueOut!=o.C001_ReturnQtyDueOut)
and o.C001_AsOfDate=#AsOfDate)
io output:
Table 'TXXX_InventoryDetail'. Scan count 0, logical reads 333952
Table 'TXXX_InventoryFull'. Scan count 1, logical reads 630
Now..the reason I think it is the if exists is that if I remove it and do a select count(*) like this:
select COUNT(*)
from dbo.T001_InventoryDetail o
where exists (select 1
from dbo.TXXX_InventoryFull i
where i.C001_AsOfDate= o.C001_AsOfDate
and i.C001_ProductID=o.C001_ProductID
and i.C001_StoreNumber=o.C001_StoreNumber
and i.C001_AsOfDate=#AsOfDate
and (i.C001_LastModelDate!=o.C001_LastModelDate
or o.C001_InventoryQty!=o.C001_InventoryQty
or i.C001_OnOrderQty!=o.C001_OnOrderQty
or i.C001_TBOQty!=o.C001_TBOQty
or i.C001_ModelQty!=o.C001_ModelQty
or i.C001_TBOAdjustQty!=o.C001_TBOAdjustQty
or i.C001_ReturnQtyPending!=o.C001_ReturnQtyPending
or i.C001_ReturnQtyInProcess!=o.C001_ReturnQtyInProcess
or i.C001_ReturnQtyDueOut!=o.C001_ReturnQtyDueOut))
and o.C001_AsOfDate=#AsOfDate
TXXX_InventoryFull'. Scan count 41, logical reads 692
T001_InventoryDetail'. Scan count 65, logical reads 17477
Worktable'. Scan count 0, logical reads 0
It is generally said that one should avoid doing coordinated subqueries in the predicate, as these tend to force nested loop joins. When querying large datasets, especially where one's trying to discover a difference between the sets, it's important to allow the query optimizer to choose dynamically between hash, merge and nested loop algorhithms, which may not be possible if the query is structured using a coordinated subquery. Better to create these as derived tables in the FROM clause.
I have found similar issues using the EXISTS statement on a SQL 08 R2 server, where the exact same statement runs fine on SQL 08 and SQL 05.
I found that changing something like
WHILE EXISTS(SELECT * FROM X)
Would be super slow, but:
WHILE ISNULL((SELECT TOP 1 ID FROM X), 0) <> 0
Runs perfectly fast again.
To me, it seems like an R2 issue...
I would guess that the plan you get is quite different when you use the join. Perhaps the imbalance in the number of rows (very large outer table, smaller inner table) is giving the optimizer fits, but it can probably eliminate rows much easier with the join (you'll probably see additional loop operators with the worse query). Tough to really guess without seeing the plans or being able to reproduce, but you should always aim at eliminating the most rows as early in the plan as possible. Pulling back millions of rows through several operators / subqueries only to eliminate most of them later in the plan is almost certainly going to yield worse performance.
Related
I need to know the number of rows in a table to calculate a percentage. If the total count is greater than some predefined constant, I will use the constant value. Otherwise, I will use the actual number of rows.
I can use SELECT count(*) FROM table. But if my constant value is 500,000 and I have 5,000,000,000 rows in my table, counting all rows will waste a lot of time.
Is it possible to stop counting as soon as my constant value is surpassed?
I need the exact number of rows only as long as it's below the given limit. Otherwise, if the count is above the limit, I use the limit value instead and want the answer as fast as possible.
Something like this:
SELECT text,count(*), percentual_calculus()
FROM token
GROUP BY text
ORDER BY count DESC;
Counting rows in big tables is known to be slow in PostgreSQL. The MVCC model requires a full count of live rows for a precise number. There are workarounds to speed this up dramatically if the count does not have to be exact like it seems to be in your case.
(Remember that even an "exact" count is potentially dead on arrival under concurrent write load.)
Exact count
Slow for big tables.
With concurrent write operations, it may be outdated the moment you get it.
SELECT count(*) AS exact_count FROM myschema.mytable;
Estimate
Extremely fast:
SELECT reltuples AS estimate FROM pg_class where relname = 'mytable';
Typically, the estimate is very close. How close, depends on whether ANALYZE or VACUUM are run enough - where "enough" is defined by the level of write activity to your table.
Safer estimate
The above ignores the possibility of multiple tables with the same name in one database - in different schemas. To account for that:
SELECT c.reltuples::bigint AS estimate
FROM pg_class c
JOIN pg_namespace n ON n.oid = c.relnamespace
WHERE c.relname = 'mytable'
AND n.nspname = 'myschema';
The cast to bigint formats the real number nicely, especially for big counts.
Better estimate
SELECT reltuples::bigint AS estimate
FROM pg_class
WHERE oid = 'myschema.mytable'::regclass;
Faster, simpler, safer, more elegant. See the manual on Object Identifier Types.
Replace 'myschema.mytable'::regclass with to_regclass('myschema.mytable') in Postgres 9.4+ to get nothing instead of an exception for invalid table names. See:
How to check if a table exists in a given schema
Better estimate yet (for very little added cost)
This does not work for partitioned tables because relpages is always -1 for the parent table (while reltuples contains an actual estimate covering all partitions) - tested in Postgres 14.
You have to add up estimates for all partitions instead.
We can do what the Postgres planner does. Quoting the Row Estimation Examples in the manual:
These numbers are current as of the last VACUUM or ANALYZE on the
table. The planner then fetches the actual current number of pages in
the table (this is a cheap operation, not requiring a table scan). If
that is different from relpages then reltuples is scaled
accordingly to arrive at a current number-of-rows estimate.
Postgres uses estimate_rel_size defined in src/backend/utils/adt/plancat.c, which also covers the corner case of no data in pg_class because the relation was never vacuumed. We can do something similar in SQL:
Minimal form
SELECT (reltuples / relpages * (pg_relation_size(oid) / 8192))::bigint
FROM pg_class
WHERE oid = 'mytable'::regclass; -- your table here
Safe and explicit
SELECT (CASE WHEN c.reltuples < 0 THEN NULL -- never vacuumed
WHEN c.relpages = 0 THEN float8 '0' -- empty table
ELSE c.reltuples / c.relpages END
* (pg_catalog.pg_relation_size(c.oid)
/ pg_catalog.current_setting('block_size')::int)
)::bigint
FROM pg_catalog.pg_class c
WHERE c.oid = 'myschema.mytable'::regclass; -- schema-qualified table here
Doesn't break with empty tables and tables that have never seen VACUUM or ANALYZE. The manual on pg_class:
If the table has never yet been vacuumed or analyzed, reltuples contains -1 indicating that the row count is unknown.
If this query returns NULL, run ANALYZE or VACUUM for the table and repeat. (Alternatively, you could estimate row width based on column types like Postgres does, but that's tedious and error-prone.)
If this query returns 0, the table seems to be empty. But I would ANALYZE to make sure. (And maybe check your autovacuum settings.)
Typically, block_size is 8192. current_setting('block_size')::int covers rare exceptions.
Table and schema qualifications make it immune to any search_path and scope.
Either way, the query consistently takes < 0.1 ms for me.
More Web resources:
The Postgres Wiki FAQ
The Postgres wiki pages for count estimates and count(*) performance
TABLESAMPLE SYSTEM (n) in Postgres 9.5+
SELECT 100 * count(*) AS estimate FROM mytable TABLESAMPLE SYSTEM (1);
Like #a_horse commented, the added clause for the SELECT command can be useful if statistics in pg_class are not current enough for some reason. For example:
No autovacuum running.
Immediately after a large INSERT / UPDATE / DELETE.
TEMPORARY tables (which are not covered by autovacuum).
This only looks at a random n % (1 in the example) selection of blocks and counts rows in it. A bigger sample increases the cost and reduces the error, your pick. Accuracy depends on more factors:
Distribution of row size. If a given block happens to hold wider than usual rows, the count is lower than usual etc.
Dead tuples or a FILLFACTOR occupy space per block. If unevenly distributed across the table, the estimate may be off.
General rounding errors.
Typically, the estimate from pg_class will be faster and more accurate.
Answer to actual question
First, I need to know the number of rows in that table, if the total
count is greater than some predefined constant,
And whether it ...
... is possible at the moment the count pass my constant value, it will
stop the counting (and not wait to finish the counting to inform the
row count is greater).
Yes. You can use a subquery with LIMIT:
SELECT count(*) FROM (SELECT 1 FROM token LIMIT 500000) t;
Postgres actually stops counting beyond the given limit, you get an exact and current count for up to n rows (500000 in the example), and n otherwise. Not nearly as fast as the estimate in pg_class, though.
I did this once in a postgres app by running:
EXPLAIN SELECT * FROM foo;
Then examining the output with a regex, or similar logic. For a simple SELECT *, the first line of output should look something like this:
Seq Scan on uids (cost=0.00..1.21 rows=8 width=75)
You can use the rows=(\d+) value as a rough estimate of the number of rows that would be returned, then only do the actual SELECT COUNT(*) if the estimate is, say, less than 1.5x your threshold (or whatever number you deem makes sense for your application).
Depending on the complexity of your query, this number may become less and less accurate. In fact, in my application, as we added joins and complex conditions, it became so inaccurate it was completely worthless, even to know how within a power of 100 how many rows we'd have returned, so we had to abandon that strategy.
But if your query is simple enough that Pg can predict within some reasonable margin of error how many rows it will return, it may work for you.
Reference taken from this Blog.
You can use below to query to find row count.
Using pg_class:
SELECT reltuples::bigint AS EstimatedCount
FROM pg_class
WHERE oid = 'public.TableName'::regclass;
Using pg_stat_user_tables:
SELECT
schemaname
,relname
,n_live_tup AS EstimatedCount
FROM pg_stat_user_tables
ORDER BY n_live_tup DESC;
How wide is the text column?
With a GROUP BY there's not much you can do to avoid a data scan (at least an index scan).
I'd recommend:
If possible, changing the schema to remove duplication of text data. This way the count will happen on a narrow foreign key field in the 'many' table.
Alternatively, creating a generated column with a HASH of the text, then GROUP BY the hash column.
Again, this is to decrease the workload (scan through a narrow column index)
Edit:
Your original question did not quite match your edit. I'm not sure if you're aware that the COUNT, when used with a GROUP BY, will return the count of items per group and not the count of items in the entire table.
You can also just SELECT MAX(id) FROM <table_name>; change id to whatever the PK of the table is
In Oracle, you could use rownum to limit the number of rows returned. I am guessing similar construct exists in other SQLs as well. So, for the example you gave, you could limit the number of rows returned to 500001 and apply a count(*) then:
SELECT (case when cnt > 500000 then 500000 else cnt end) myCnt
FROM (SELECT count(*) cnt FROM table WHERE rownum<=500001)
For SQL Server (2005 or above) a quick and reliable method is:
SELECT SUM (row_count)
FROM sys.dm_db_partition_stats
WHERE object_id=OBJECT_ID('MyTableName')
AND (index_id=0 or index_id=1);
Details about sys.dm_db_partition_stats are explained in MSDN
The query adds rows from all parts of a (possibly) partitioned table.
index_id=0 is an unordered table (Heap) and index_id=1 is an ordered table (clustered index)
Even faster (but unreliable) methods are detailed here.
I have a requirement to transfer data from 2 tables (Table A and Table B) into a new table.
I am using a query to join both A and B tables using an ID column.
Table A and B are archive tables without any indexes. (Millions of records)
Table X and Y are a replica of A and B with good indexes. (Some thousands of records)
Below is the code for my project.
with data as
(
SELECT a.*, b.* FROM A_archive a
join B_archive b where a.transaction_id = b.transaction_id
UNION
SELECT x.*, y.* FROM X x
join Y y where x.transaction_id = y.transaction_id
)
INSERT INTO
Another_Table
(
columns
)
select * from data
On Conflict(transaction_id)
do udpate ...
The above whole thing is running in production environment and has nearly 140 million records.
Due to this production database is taking almost 10 hours to process the data and failing.
I am also having a distributed job scheduler in AWS to schedule this query inside a function and retrieve the latest records every 5 hours. The archive tables store closed invoice data. Pega UI will be using this table for retrieving data about closed invoices and showing to the customer.
Please suggest something that is a bit more performant.
UNION removes duplicate rows. On big unindexed tables that is an expensive operation. Try UNION ALL if you don't need deduplication. It will save the s**tton of data shuffling and comparisons required for deduplication.
Without indexes on your archival tables your JOIN operation will be grossly inefficient. Index, at a minimum, the transaction_id columns you use in your ON clause.
You don't say what you want to do with the resulting table. In many cases you'll be able to use a VIEW rather than a table for your purposes. A VIEW removes the work of creating the derived table. Actually it defers the work to the time of SELECT operations using the derived structure. If your SELECT operations have highly selective WHERE clauses the savings can be astonishing. For this to work well you may need to put appropriate indexes on your archival tables.
You use SELECT * when you could enumerate the columns you need. That certainly puts one redundant column into your result: it generates two copies of transaction_id. It also may generate other redundant or unused data. Always avoid SELECT * in production software unless you know you need it.
Keep this in mind: SQL is declarative, not procedural. You declare (describe) the result you require, and you let the server work out the best way to get it. VIEWs let the server do this work for you in cases like your table combination. It will use the indexes you provide as best it can.
That UNION must be costly, it pretty much builds a temp-table in the background containing all the A-B + X-Y records, sorts it (over all fields) and then removes any doubles. If you say 100 million records are involved then that's a LOT of sorting going on that most likely will involve swapping out to disk.
Keep in mind that you only need to do this if there are expected duplicates
in the result from the JOIN between A and B
in the result from the JOIN between X and Y
in the combined result from the two above
IF neither of those are expected, just use UNION ALL
In fact, in that case, why not have 1 INSERT operation for A-B and another one for X-Y? Going by the description I'd say that whatever is in X-Y should overrule whatever is in A-B anyway, right?
Also, as mentioned by O.Jones, archive tables or not, they should come at least with a (preferably clustered) index on the transaction_id fields you're JOINing on. (same for the Another_Table btw)
All that said, processing 100M records in 1 transaction IS going to take some time, it's just a lot of data that's being moved around. But 10h does sound excessive indeed.
I am trying to optimize a query which has been destroying my DB.
https://explain.depesz.com/s/isM1
If you have any insights into how to make this better please let me know.
We are using RDS/Postgres 11.9
explain analyze SELECT "src_rowdifference"."key",
"src_rowdifference"."port_id",
"src_rowdifference"."shipping_line_id",
"src_rowdifference"."container_type_id",
"src_rowdifference"."shift_id",
"src_rowdifference"."prev_availability_id",
"src_rowdifference"."new_availability_id",
"src_rowdifference"."date",
"src_rowdifference"."prev_last_update",
"src_rowdifference"."new_last_update"
FROM "src_rowdifference"
INNER JOIN "src_containertype" ON ("src_rowdifference"."container_type_id" = "src_containertype"."key")
WHERE ("src_rowdifference"."container_type_id" IN
(SELECT U0."key"
FROM "src_containertype" U0
INNER JOIN "notification_tablenotification_container_types" U1 ON (U0."key" = U1."containertype_id")
WHERE U1."tablenotification_id" = 'test#test.com')
AND "src_rowdifference"."new_last_update" >= '2020-01-15T03:11:06.291947+00:00'::timestamptz
AND "src_rowdifference"."port_id" IN
(SELECT U0."key"
FROM "src_port" U0
INNER JOIN "notification_tablenotification_ports" U1 ON (U0."key" = U1."port_id")
WHERE U1."tablenotification_id" = 'test#test.com')
AND "src_rowdifference"."shipping_line_id" IN
(SELECT U0."key"
FROM "src_shippingline" U0
INNER JOIN "notification_tablenotification_shipping_lines" U1 ON (U0."key" = U1."shippingline_id")
WHERE U1."tablenotification_id" = 'test#test.com')
AND "src_rowdifference"."prev_last_update" IS NOT NULL
AND NOT ("src_rowdifference"."prev_availability_id" = 'na'
AND "src_rowdifference"."prev_availability_id" IS NOT NULL)
AND NOT ("src_rowdifference"."key" IN
(SELECT V1."rowdifference_id"
FROM "notification_tablenotificationtrigger_row_differences" V1
WHERE V1."tablenotificationtrigger_id" IN
(SELECT U0."id"
FROM "notification_tablenotificationtrigger" U0
WHERE U0."notification_id" = 'test#test.com'))));
All my indexes are btree + btree(varchar_pattern_ops)
"src_rowdifference_port_id_shipping_line_id_9b3465fc_uniq" UNIQUE CONSTRAINT, btree (port_id, shipping_line_id, container_type_id, shift_id, date, new_last_update)
Edit: A little unrelated change that I made was added some more ssd disk space to my RDS instance. That made a huge difference to the CPU usage and in turn made a huge difference to the number of connections we have.
It is hard to think about the plan as a whole, as I don't understand what it is looking for. But looking at the individual pieces, there are two which together dominate the run time.
One is the index scan on src_rowdifference_port_id_shipping_line_id_9b3465fc, which seems pretty slow given the number of rows returned. Comparing the Index Condition to the index columns, I can see that the condition on new_last_update cannot be applied efficiently in the index because two columns in the index come before it and have no equality conditions in the node. So instead that >= is applied as an "in-index filter" where it needs to test each row and reject it, rather than just skipping it in bulk. I don't know how many rows that removes as the "Rows Removed by Filter" does not count in-index filters, but it is potentially large. So one thing to try would be to make a new index on (port_id, shipping_line_id, container_type_id, new_last_update). Or maybe replace that index with a reordered version (port_id, shipping_line_id, container_type_id, new_last_update, shift_id, date) but of course that might make some other query worse.
The other time consuming thing is kicking the materialized node 47 thousand times (each one looping over up to 22 thousand rows) to implement NOT (SubPlan 1). That should be using a hashed subplan, rather than a linear searched subplan. The only reason I can think of that it not doing the hashed subplan is that work_mem is not large enough to anticipate fitting it into memory. What is your setting for work_mem? What happens if you bump it up to "100MB" or so?
The NOT (SubPlan 1) from the EXPLAIN corresponds to the part of your query AND NOT ("src_rowdifference"."key" IN (...)). If bumping up work_mem doesn't work, you could try rewriting that into a NOT EXISTS clause instead.
I am running two sql queries say,
select obname from table1 where obid = 12
select modname from table2 where modid = 12
Both are taking very less time, say 300 ms each.
But when I am running:
select obname, modname
from (select obname from table1 where obid = 12) as alias1,
(select modname from table2 where modid = 12) as alias2
It is taking 3500ms. Why is it so?
In general, putting two scalar queries in the from clause is not going to affect performance. In fact, from an application perspective, one query may be faster because there is less overhead going back and forth to the database. A scalar query returns one column and one row.
However, if the queries are returning multiple rows, then your little comma is doing a massive Cartesian product (which is why I always use CROSS JOIN rather than a comma in a FROM clause). In that case, all bets are off, because the data has to be processed after the results start returning.
I have a query on a postgresql 9.2 system that takes about 20s in it's normal form but only takes ~120ms when using a CTE.
I simplified both queries for brevity.
Here is the normal form (takes about 20s):
SELECT *
FROM tableA
WHERE (columna = 1 OR columnb = 2) AND
atype = 35 AND
aid IN (1, 2, 3)
ORDER BY modified_at DESC
LIMIT 25;
Here is the explain for this query: http://explain.depesz.com/s/2v8
The CTE form (about 120ms):
WITH raw AS (
SELECT *
FROM tableA
WHERE (columna = 1 OR columnb = 2) AND
atype = 35 AND
aid IN (1, 2, 3)
)
SELECT *
FROM raw
ORDER BY modified_at DESC
LIMIT 25;
Here is the explain for the CTE: http://explain.depesz.com/s/uxy
Simply by moving the ORDER BY to the outer part of the query reduces the cost by 99%.
I have two questions: 1) is there a way to construct the first query without using a CTE in such a way that it is logically equivalent more performant and 2) what does this difference in performance say about how the planner is determining how to fetch the data?
Regarding the questions above, are there additional statistics or other planner hints that would help improve the performance of the first query?
Edit: Taking away the limit also causes the query to use a heap scan as opposed to an index scan backwards. Without the LIMIT the query completes in 40ms.
After seeing the effect of the LIMIT I tried with LIMIT 1, LIMIT 2, etc. The query performs in under 100ms when using LIMIT 1 and 10s+ with LIMIT > 1.
After thinking about this some more, question 2 boils down to why does the planner use an index scan backwards in one case and a bitmap heap scan + sort in another logically equivalent case? And how can I "help" the planner use the efficient plan in both cases?
Update:
I accepted Craig's answer because it was the most comprehensive and helpful. The way I ended up solving the problem was by using a query that was practically equivalent though not logically equivalent. At the root of the issue was an index scan backwards of the index on modified_at. In order to inform the planner that this was not a good idea I add a predicate of the form WHERE modified_at >= NOW() - INTERVAL '1 year'. This included enough data for the application but prevented the planner from going down the backwards index scan path.
This was a much lower impact solution that prevented the need to rewrite the queries using either a sub query or a CTE. YMMV.
Here's why this is happening, with the following explanation current until at least 9.3 (if you're reading this and on a newer version, check to make sure it hasn't changed):
PostgreSQL doesn't optimize across CTE boundaries. Each CTE clause is run in isolation and its results are consumed by other parts of the query. So a query like:
WITH blah AS (
SELECT * FROM some_table
)
SELECT *
FROM blah
WHERE id = 4;
will cause the full inner query to get executed. PostgreSQL won't "push down" the id = 4 qualification into the inner query. CTEs are "optimization fences" in that regard, which can be both good or bad; it lets you override the planner when you want to, but prevents you from using CTEs as simple syntactic cleanup for a deeply nested FROM subquery chain if you do need push-down.
If you rephrase the above as:
SELECT *
FROM (SELECT * FROM some_table) AS blah
WHERE id = 4;
using a sub-query in FROM instead of a CTE, Pg will push the qual down into the subquery and it'll all run nice and quickly.
As you have discovered, this can also work to your benefit when the query planner makes a poor decision. It appears that in your case a backward index scan of the table is immensely more expensive a bitmap or index scan of two smaller indexes followed by a filter and sort, but the planner doesn't think it will be so it plans the query to scan the index.
When you use the CTE, it can't push the ORDER BY into the inner query, so you're overriding its plan and forcing it to use what it thinks is an inferior execution plan - but one that turns out to be much better.
There's a nasty workaround that can be used for these situations called the OFFSET 0 hack, but you should only use it if you can't figure out a way to make the planner do the right thing - and if you have to use it, please boil this down to a self-contained test case and report it to the PostgreSQL mailing list as a possible query planner bug.
Instead, I recommend first looking at why the planner is making the wrong decision.
The first candidate is stats / estimates problems, and sure enough when we look at your problematic query plan there's a factor of 3500 mis-estimation of the expected result rows. That's big, but not impossibly big, though it's more interesting that you actually only get one row where the planner is expecting a non-trivial row set. That doesn't help us much, though; if the row count is lower than expected that means that choosing to use the index was a better choice than expected.
The main issue looks like it's not using the smaller, more selective indexes sierra_kilo and papa_lima because it sees the ORDER BY and thinks that it'll save more time doing a backward index scan and avoiding the sort than it really does. That makes sense given that there's only one matching row to sort! If it got the expected 3500 rows then it might've made more sense to avoid the sort, though that's still a fairly small rowset to just sort in memory.
Do you set any parameters like enable_seqscan, etc? If you do, unset them; they're for testing only and totally inappropriate for production use. If you aren't using the enable_ params I think it's worth raising this on the PostgreSQL mailing list pgsql-perform. The anonymized plans make this a bit difficult, though, especially since there's no gurantee that identifiers from one plan refer to the same objects in the other plan, and they don't match what you wrote in the query on the question. You'll want to produce a properly hand-done version where everything matches up before asking on the mailing list.
There's a fairly good chance that you'll need to provide the real values for anyone to help. If you don't want to do that on a public mailing list, there's another option available. (I should note that I work for one of them, per my profile).
Just a shot in the dark, but what happens if you run this
SELECT *
FROM (
SELECT *
FROM tableA
WHERE (columna = 1 OR columnb = 2) AND
atype = 35 AND
aid IN (1, 2, 3)
) AS x
ORDER BY modified_at DESC
LIMIT 25;