Thread: Sorted union
Using PostgreSQL 8.0.4. I've got a table with 4.5 million rows that I expect to become huge (hundred million? who knows). Each row has a start and end time. I want to retrieve all the events during a timespan in one list; typical timespans will involve up to a couple rows. If the start and stop happen at the same time (which is possible), the start must come first in my sequence. So essentially, I want this: select when_stopped as when_happened, 1 as order_hint from transaction t where '2005-10-25 15:00:00' <= when_stopped and when_stopped <= '2005-10-26 10:00:00' union all select when_stopped as when_happened, 2 as order_hint from transaction t where '2005-10-25 15:00:00' <= when_stopped and when_stopped <= '2005-10-26 10:00:00' order by when_happened, order_hint; I'd really like the first row to be retrieved in O(1) time and the last in O(n) time (n = number of rows in the timespan, not the whole table). I previously was doing things manually with flat files. But there's a sort in PostgreSQL's plan, so I think I'm getting O(n log n) time for both. It's frustrating to start using a real database and get performance regressions. QUERY PLAN ------------------------------------------------------------------------ ------------------------------------------------------------------------ --------------------------------- Sort (cost=667469.90..676207.19 rows=3494916 width=8) (actual time=28503.612..31377.254 rows=3364006 loops=1) Sort Key: when_happened, order_hint -> Append (cost=0.00..194524.95 rows=3494916 width=8) (actual time=0.191..14659.712 rows=3364006 loops=1) -> Subquery Scan "*SELECT* 1" (cost=0.00..97262.48 rows=1747458 width=8) (actual time=0.190..5375.925 rows=1682003 loops=1) -> Index Scan using transaction_stopped on "transaction" t (cost=0.00..79787.90 rows=1747458 width=8) (actual time=0.186..2962.585 rows=1682003 loops=1) Index Cond: (('2005-10-25 15:00:00'::timestamp without time zone <= when_stopped) AND (when_stopped <= '2005-10-26 10:00:00'::timestamp without time zone)) -> Subquery Scan "*SELECT* 2" (cost=0.00..97262.48 rows=1747458 width=8) (actual time=0.167..5449.151 rows=1682003 loops=1) -> Index Scan using transaction_stopped on "transaction" t (cost=0.00..79787.90 rows=1747458 width=8) (actual time=0.163..3026.730 rows=1682003 loops=1) Index Cond: (('2005-10-25 15:00:00'::timestamp without time zone <= when_stopped) AND (when_stopped <= '2005-10-26 10:00:00'::timestamp without time zone)) Total runtime: 33312.814 ms (10 rows) Each side of the union is retrieved in sorted order, but it doesn't seem to realize that. There seem to be two things it's missing: (1) It doesn't recognize that constant expressions are irrelevant to the sort. I.e., the first half of the union: select when_started as when_happened, 1 as order_hint from transaction t where '2005-10-25 15:00:00' <= when_started and when_started <= '2005-10-26 10:00:00' order by when_happened, order_hint; does this: QUERY PLAN ------------------------------------------------------------------------ ------------------------------------------------------------------------ --------------------- Sort (cost=291770.42..296139.05 rows=1747453 width=8) (actual time=8462.026..9895.715 rows=1681994 loops=1) Sort Key: when_started, 1 -> Index Scan using transaction_started on "transaction" t (cost=0.00..79788.21 rows=1747453 width=8) (actual time=0.190..2953.393 rows=1681994 loops=1) Index Cond: (('2005-10-25 15:00:00'::timestamp without time zone <= when_started) AND (when_started <= '2005-10-26 10:00:00'::timestamp without time zone)) Total runtime: 10835.114 ms (5 rows) The sort is unnecessary. If I take out the constant order_hint, it works: select when_started as when_happened from transaction t where '2005-10-25 15:00:00' <= when_started and when_started <= '2005-10-26 10:00:00' order by when_happened; QUERY PLAN ------------------------------------------------------------------------ ------------------------------------------------------------------------ --------------- Index Scan using transaction_started on "transaction" t (cost=0.00..79788.21 rows=1747453 width=8) (actual time=0.189..2715.513 rows=1681994 loops=1) Index Cond: (('2005-10-25 15:00:00'::timestamp without time zone <= when_started) AND (when_started <= '2005-10-26 10:00:00'::timestamp without time zone)) Total runtime: 3630.817 ms (3 rows) (2) It doesn't recognize that each half of the union is sorted and thus they only need to be merged. This is true even without the order_hint bits: select when_stopped as when_happened from transaction t where '2005-10-25 15:00:00' <= when_stopped and when_stopped <= '2005-10-26 10:00:00' union all select when_stopped as when_happened from transaction t where '2005-10-25 15:00:00' <= when_stopped and when_stopped <= '2005-10-26 10:00:00' order by when_happened; QUERY PLAN ------------------------------------------------------------------------ ------------------------------------------------------------------------ --------------------------------- Sort (cost=667469.90..676207.19 rows=3494916 width=8) (actual time=28088.783..30898.854 rows=3364006 loops=1) Sort Key: when_happened -> Append (cost=0.00..194524.95 rows=3494916 width=8) (actual time=0.153..14410.485 rows=3364006 loops=1) -> Subquery Scan "*SELECT* 1" (cost=0.00..97262.48 rows=1747458 width=8) (actual time=0.152..5287.092 rows=1682003 loops=1) -> Index Scan using transaction_stopped on "transaction" t (cost=0.00..79787.90 rows=1747458 width=8) (actual time=0.149..2885.905 rows=1682003 loops=1) Index Cond: (('2005-10-25 15:00:00'::timestamp without time zone <= when_stopped) AND (when_stopped <= '2005-10-26 10:00:00'::timestamp without time zone)) -> Subquery Scan "*SELECT* 2" (cost=0.00..97262.48 rows=1747458 width=8) (actual time=0.152..5254.425 rows=1682003 loops=1) -> Index Scan using transaction_stopped on "transaction" t (cost=0.00..79787.90 rows=1747458 width=8) (actual time=0.149..2905.861 rows=1682003 loops=1) Index Cond: (('2005-10-25 15:00:00'::timestamp without time zone <= when_stopped) AND (when_stopped <= '2005-10-26 10:00:00'::timestamp without time zone)) Total runtime: 32766.566 ms (10 rows) Is there some way I can work around this? The best I can think of now is to open two connections, one for each half of the union. I can do the merge manually on the client side. It'd work, but I'd really prefer the database server take care of this for me. -- Scott Lamb <http://www.slamb.org/>
On 2 Nov 2005, at 21:13, Scott Lamb wrote: > I want to retrieve all the events during a timespan in one list; > typical timespans will involve up to a couple rows. Err, I meant up to a couple million rows. With two rows, I wouldn't be so concerned about performance. ;) -- Scott Lamb <http://www.slamb.org/>
> select when_stopped as when_happened, > 1 as order_hint > from transaction t > where '2005-10-25 15:00:00' <= when_stopped > and when_stopped <= '2005-10-26 10:00:00' > union all > select when_stopped as when_happened, > 2 as order_hint > from transaction t > where '2005-10-25 15:00:00' <= when_stopped > and when_stopped <= '2005-10-26 10:00:00' > order by when_happened, order_hint; hmm, try pushing the union into a subquery...this is better style because it's kind of ambiguous if the ordering will apply before/after the union. select q.when from ( select 1 as hint, start_time as when [...] union all select 2 as hint, end_time as when [...] ) q order by q.seq, when question: why do you want to flatten the table...is it not easier to work with as records? Merlin
Merlin Moncure wrote: > hmm, try pushing the union into a subquery...this is better style > because it's kind of ambiguous if the ordering will apply before/after > the union. Seems to be a little slower. There's a new "subquery scan" step. explain analyze select q.when_happened from ( select when_stopped as when_happened, 1 as order_hint from transaction t where '2005-10-25 15:00:00' <= when_stopped and when_stopped <= '2005-10-26 10:00:00' union all select when_stopped as when_happened, 2 as order_hint from transaction t where '2005-10-25 15:00:00' <= when_stopped and when_stopped <= '2005-10-26 10:00:00' ) q order by when_happened, order_hint; QUERY PLAN --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Sort (cost=713013.96..721751.25 rows=3494916 width=12) (actual time=34392.264..37237.148 rows=3364006 loops=1) Sort Key: when_happened, order_hint -> Subquery Scan q (cost=0.00..229474.11 rows=3494916 width=12) (actual time=0.194..20283.452 rows=3364006 loops=1) -> Append (cost=0.00..194524.95 rows=3494916 width=8) (actual time=0.191..14967.632 rows=3364006 loops=1) -> Subquery Scan "*SELECT* 1" (cost=0.00..97262.48 rows=1747458 width=8) (actual time=0.189..5535.139 rows=1682003 loops=1) -> Index Scan using transaction_stopped on "transaction" t (cost=0.00..79787.90 rows=1747458 width=8) (actual time=0.186..3097.268 rows=1682003 loops=1) Index Cond: (('2005-10-25 15:00:00'::timestamp without time zone <= when_stopped) AND (when_stopped <= '2005-10-26 10:00:00'::timestamp without time zone)) -> Subquery Scan "*SELECT* 2" (cost=0.00..97262.48 rows=1747458 width=8) (actual time=0.173..5625.155 rows=1682003 loops=1) -> Index Scan using transaction_stopped on "transaction" t (cost=0.00..79787.90 rows=1747458 width=8) (actual time=0.169..3146.714 rows=1682003 loops=1) Index Cond: (('2005-10-25 15:00:00'::timestamp without time zone <= when_stopped) AND (when_stopped <= '2005-10-26 10:00:00'::timestamp without time zone)) Total runtime: 39775.225 ms (11 rows) > question: why do you want to flatten the table...is it not easier to > work with as records? For most things, yes. But I'm making a bunch of different graphs from these data, and a few of them are much easier with events. The best example is my concurrency graph. Whenever there's a start event, it goes up one. Whenever there's a stop event, it goes down one. It's completely trivial once you have it separated into events. Thanks, Scott
> Merlin Moncure wrote: > > hmm, try pushing the union into a subquery...this is better style > > because it's kind of ambiguous if the ordering will apply before/after > > the union. > > Seems to be a little slower. There's a new "subquery scan" step. I figured. However it's more correct, I'm not sure if the original query is necessarily guaranteed to give the right answer (in terms of ordering). It might though. > > > question: why do you want to flatten the table...is it not easier to > > work with as records? > > For most things, yes. But I'm making a bunch of different graphs from > these data, and a few of them are much easier with events. The best > example is my concurrency graph. Whenever there's a start event, it goes > up one. Whenever there's a stop event, it goes down one. It's completely > trivial once you have it separated into events. well, if you don't mind attempting things that are not trivial, how about trying: select t, (select count(*) from transaction where t between happened and when_stopped) from ( select ((generate_series(1,60) * scale)::text::interval) + '12:00 pm'::time as t ) q; for example, to check concurrency at every second for a minute (starting from 1 second) after 12:00 pm, (scale is zero in this case), select t, (select count(*) from transaction where t between happened and when_stopped) from ( select (generate_series(1,60)::text::interval) + '12:00 pm'::time as t ) q; this could be a win depending on how much data you pull into your concurrency graph. maybe not though. Merlin
On Nov 3, 2005, at 8:20 AM, Merlin Moncure wrote: > select t, (select count(*) from transaction where t between happened > and when_stopped) from > ( > select ((generate_series(1,60) * scale)::text::interval) + '12:00 > pm'::time as t > ) q; Wow. I hadn't known about generate_series, but there are a bunch of places I've needed it. As cool as this is, though, I don't think it helps me. There's another event-driven graph that I need. For lack of a better name, I call it the slot graph. Every single transaction is graphed as a horizontal line from its start time to its end time, with a vertical line at the start and stop. Successful, timed out, and failed transactions are green, black, and red, respectively. I use it in a couple different ways: (1) on short timescales, it's nice to look at individual transactions. My tester will max out at either a rate or a concurrency. If I'm having problems, I'll get bursts of timeouts. This graph is the one that makes it clear why - it shows how things align, etc. Actually, even for longer timespans, this is still helpful - it's nice to see that most of the slots are filled with timing-out transactions when the rate falls. (2) It can show you if something affects all of the transactions at once. When we did a database failover test, we saw a bunch of failures (as expected; our application isn't responsible for retries). This graph is the one that showed us that _all_ transactions that were active at a specific time failed and that no other transactions failed. (There was a sharp vertical line of reds and blacks in the larger block of greens). I wish I could just show these to you, rather than describing them. It's all proprietary data, though. Maybe soon I'll have similar graphs of my open source SSL proxy. But the point is, I don't think I can represent this information without sending every data point to my application. I assign slots by the start time and free them by the stop time. But I think there is something I can do: I can just do a query of the transaction table sorted by start time. My graph tool can keep a priority queue of all active transactions, keyed by the stop time. Whenever it grabs a new event, it can peek at the next start time but check if there are any stop times before it. Then at the end, it can pick up the rest of the stop times. The concurrency will never exceed a few thousand, so the additional CPU time and memory complexity are not a problem. As a bonus, I will no longer need my index on the stop time. Dropping it will save a lot of disk space. Thanks for getting me off the "I need a fast query that returns these exact results" mindset. It is good to step back and look at the big picture. Mind you, I still think PostgreSQL should be able to perform that sorted union fast. Maybe sometime I'll have enough free time to take my first plunge into looking at a database query planner. Regards, Scott -- Scott Lamb <http://www.slamb.org/>
> Wow. I hadn't known about generate_series, but there are a bunch of > places I've needed it. It's a wonder tool :). > But I think there is something I can do: I can just do a query of the > transaction table sorted by start time. My graph tool can keep a Reading the previous paragraphs I was just about to suggest this. This is a much more elegant method...you are reaping the benefits of having normalized your working set. You were trying to denormalize it back to what you were used to. Yes, now you can drop your index and simplify your queries...normalized data is always more 'natural'. > Mind you, I still think PostgreSQL should be able to perform that > sorted union fast. Maybe sometime I'll have enough free time to take > my first plunge into looking at a database query planner. I'm not so sure I agree, by using union you were basically pulling two independent sets (even if they were from the same table) that needed to be ordered. There is zero chance of using the index here for ordering because you are ordering a different set than the one being indexed. Had I not been able to talk you out of de-normalizing your table I was going to suggest rigging up a materialized view and indexing that: http://jonathangardner.net/PostgreSQL/materialized_views/matviews.html Merlin
The ANSI/ISO specs are not at all ambiguous on this. An ORDER BY is not allowed for the SELECT statements within a UNION. It must come at the end and applied to the resulting UNION. Similarly, the column names in the result come from the first query in the UNION. Column names in the query on the right side of a UNION are immaterial. Unless we have reason to believe that PostgreSQL is non-compliant on this point, I don't think it is a good idea to slow the query down with the subquery. -Kevin >>> "Merlin Moncure" <merlin.moncure@rcsonline.com> >>> > Merlin Moncure wrote: > > hmm, try pushing the union into a subquery...this is better style > > because it's kind of ambiguous if the ordering will apply before/after > > the union. > > Seems to be a little slower. There's a new "subquery scan" step. I figured. However it's more correct, I'm not sure if the original query is necessarily guaranteed to give the right answer (in terms of ordering). It might though.
> The ANSI/ISO specs are not at all ambiguous on this. An > ORDER BY is not allowed for the SELECT statements within > a UNION. It must come at the end and applied to the resulting > UNION. Interesting :/ Merlin
On Nov 3, 2005, at 10:21 AM, Merlin Moncure wrote: > Reading the previous paragraphs I was just about to suggest this. > This > is a much more elegant method...you are reaping the benefits of having > normalized your working set. You were trying to denormalize it > back to > what you were used to. Yes, now you can drop your index and simplify > your queries...normalized data is always more 'natural'. I'm not sure normalized is the right word. In either case, I'm storing it in the same form. In either case, my ConcurrencyProcessor class gets the same form. The only difference is if the database splits the rows or if my application does so. But we're essentially agreed. This is the algorithm I'm going to try implementing, and I think it will work out well. It also means sending about half as much data from the database to the application. >> Mind you, I still think PostgreSQL should be able to perform that >> sorted union fast. Maybe sometime I'll have enough free time to take >> my first plunge into looking at a database query planner. > > I'm not so sure I agree, by using union you were basically pulling two > independent sets (even if they were from the same table) that > needed to > be ordered. Yes. > There is zero chance of using the index here for ordering > because you are ordering a different set than the one being indexed. I don't think that's true. It just needs to look at the idea of independently ordering each element of the union and then merging that, compared to the cost of grabbing the union and then ordering it. In this case, the former cost is about 0 - it already has independently ordered them, and the merge algorithm is trivial. <http://en.wikipedia.org/wiki/Merge_algorithm> Regards, Scott -- Scott Lamb <http://www.slamb.org/>
Just as an FYI, if you want to reassure yourself that the ORDER BY is being applied as intended, you could do the following: ( select 1 as hint, start_time as when [...] union all select 2 as hint, end_time as when [...] ) order by seq, when This is ANSI/ISO standard, and works in PostgreSQL (based on a quick test). >>> "Merlin Moncure" <merlin.moncure@rcsonline.com> >>> hmm, try pushing the union into a subquery...this is better style because it's kind of ambiguous if the ordering will apply before/after the union. select q.when from ( select 1 as hint, start_time as when [...] union all select 2 as hint, end_time as when [...] ) q order by q.seq, when