Este artigo est� escrito em Ingl�s e Portugu�s
English version.
In this article I'll show you a curious situation related to the Informix query optimizer. There are several reasons to do this. First, this is a curious behavior which by itself deserves a few lines. Another reason is that I truly admire the people who write this stuff... I did and I still do some programming, but my previous experience was related to business applications and currently it's mostly scripts, small utilities, stored procedures etc. I don't mean to offend anyone, but I honestly believe these two programming worlds are on completely different leagues. All this to say that a query optimizer is something admirable given the complexity of the issues it has to solve.
I also have an empiric view or impression regarding several query optimizers in the database market. This impression comes from my personal experience, my contact with professionals who work with other technologies and some readings... Basically I think one of our biggest competitors has an optimizer with inconstant behavior, difficult to understand and not very trustworthy. One of the other IBM relational databases, in this case DB2, has probably the most "intelligent" optimizer, but I have nearly no experience with it. As for Informix, I consider it not the most sophisticated (lot of improvements in V10 and V11, and more to come in Panther), but I really love it for being robust. Typically it just needs statistics. I think I only used or recommended directives two times since I work with Informix. And I believe I never personally hit an optimizer bug (and if you check the fix lists in each fixpack you'll see some of them there...)
To end the introduction I'd like to add that one of my usual tasks is query optimizing... So I did catch a few interesting situations. As I wrote above, typically an UPDATE statistics solves the problem. A few times I saw situations where I noticed optimizer limitations (things could work better if it could solve the query another way), or even engine limitations (the query plan is good, but the execution takes unnecessary time to complete). The following situation was brought to my attention by a customer who was looking around queries using a specific table which had a reasonable number of sequential scans.... Let's see the table schema first:
create table example_table
(
et_id serial not null ,
et_ts datetime year to second,
et_col1 smallint,
et_col2 integer,
et_var1 varchar(50),
et_var2 varchar(50),
et_var3 varchar(50),
et_var4 varchar(50),
et_var5 varchar(50),
primary key (et_id)
)
lock mode row;
create index ids_example_table_1 on example_table (et_timestamp) using btree in dbs1
And then two query plans they manage to obtain:
QUERY: (OPTIMIZATION TIMESTAMP: 08-25-2010 12:31:56)
------
SELECT et_ts, et_col2
FROM example_table WHERE et_var2='LessCommonValue'
AND et_col1 = 1 ORDER BY et_ts DESC
Estimated Cost: 5439
Estimated # of Rows Returned: 684
Temporary Files Required For: Order By
1) informix.example_table: SEQUENTIAL SCAN
Filters: (informix.example_table.et_var2 = 'LessCommonValue' AND informix.example_table.et_col1 = 1 )
So this was the query responsible for their sequential scans. Nothing special. The conditions are on non indexed fields. So no index is used. It includes an ORDER BY clause, so a temporary file to sort is needed (actually it can be done in memory, but the plan means that a sort operation has to be done).
So, nothing noticeable here... But let's check another query plan on the same table. In fact, the query is almost the same:
QUERY: (OPTIMIZATION TIMESTAMP: 08-25-2010 12:28:32)
------
SELECT et_ts, et_col2
FROM example_table WHERE et_var2='MoreCommonValue'
AND et_col1 = 1 ORDER BY et_ts DESC
Estimated Cost: 5915
Estimated # of Rows Returned: 2384
1) informix.example_table: INDEX PATH
Filters: (informix.example_table.et_var2 = 'MoreCommonValue' AND informix.example_table.et_col1 = 1 )
(1) Index Name: informix.ids_example_table_1
Index Keys: et_ts (Serial, fragments: ALL)
Ok... It's almost the same query, but this time we have an INDEX path.... Wait... Didn't I wrote that the conditions where not on indexed column(s)? Yes. Another difference: It does not need a temporary file for sort (meaning that it's not doing a sort operation). Also note that there are no filter applied on the index. So, in this case it's doing a completely different thing than on the previous situation. It reads the index (the whole index), accesses the rows (so it gets them already ordered) and then applies the filters discarding the rows that do not match. It does that because it wants to avoid the sort operation. But why does it change the behavior? You might have noticed that I already included a clue about this. To protect data privacy, I changed the table name, table fields and the query values. And in the first query I used "LessCommonValue" and on the second I used "MoreCommonValue". So as you might expect, the value on the second query is more frequent in the table than the value of the first query. How do we know that? First, by looking at the number of estimated rows returned (684 versus 2384). How does it predict this values? Because the table has statistics, and in particular column distribution. Here are them for the et_var2 column (taken with dbschema -hd):
Distribution for informix.example_table.et_var2
Constructed on 2010-08-25 12:19:04.31216
Medium Mode, 2.500000 Resolution, 0.950000 Confidence
--- DISTRIBUTION ---
( AnotherValue )
--- OVERFLOW ---
1: ( 4347, Value_1 )
2: ( 4231, Value_2 )
3: ( 3129, MoreCommonValue )
4: ( 2840, Value_4 )
5: ( 2405, Value_5 )
6: ( 3854, Value_6 )
7: ( 3941, Value_7 )
8: ( 2086, Value_8 )
9: ( 4086, Value_9 )
10: ( 4144, Value_10 )
11: ( 4086, Value_11 )
12: ( 2666, Value_12 )
13: ( 2869, Value_13 )
14: ( 3245, Value_14 )
15: ( 2811, Value_15 )
16: ( 3419, Value_16 )
17: ( 2637, Value_17 )
18: ( 3187, Value_18 )
19: ( 4347, Value_19 )
20: ( 898, LessCommonValue )
21: ( 4144, Value_20 )
22: ( 2260, Value_21 )
23: ( 3999, Value_22 )
24: ( 1797, Value_23 )
25: ( 2115, Value_24 )
26: ( 2173, Value_25 )
27: ( 4144, Value_26 )
If you're not used to look at this output I'll describe it. The distributions are showed as a list of "bins". Each bin, except the first one, contains 3 columns: The number of record it represents, the number of unique values within that bin, and the highest value within that bin. The above is not a good example because the "normal" bin values has only one (AnotherValue) and given it's the first one it means it's the "lowest" value in the column. Let's see an example from the dbschema description in the migration guide:
( 5)
1: ( 16, 7, 11)
2: ( 16, 6, 17)
3: ( 16, 8, 25)
4: ( 16, 8, 38)
5: ( 16, 7, 52)
6: ( 16, 8, 73)
7: ( 16, 12, 95)
8: ( 16, 12, 139)
9: ( 16, 11, 182)
10: ( 10, 5, 200)
Ok. So, looking at this we could see that the lowest value is 5, and that between 5 and 11 we have 16 values, 7 of them are unique. Between 11 and 17 we have 6 unique values etc.
Going back to our example we see an "OVERFLOW" section. This is created when there are highly repeated values that would skew the distributions. In these cases the repeated values are showed along the number of times they appear.
So, in our case, the values we're interested are "LessCommonValue" (898 times) and "MoreCommonValue" (3129 times). So with these evidences we can understand the decision to choose the INDEX PATH. When the engine gets to the end of the result set it's already ordered. And in this situation (more than 3000 rows) the order by would be much more expensive than for the other value (around 800 rows).
It's very arguable if the choice is effectively correct. But my point is that it cares... Meaning it goes deep in it's analysis, and that it takes into account many aspects, and is able to decide to use a query plan that it's not obvious (and choose it for good reasons).
Vers�o portuguesa:
Neste artigo vou mostrar uma situa��o curiosa relacionada com o optimizador de queries do Informix. H� v�rias raz�es para fazer isto. Primeiro porque � um comportamento interessante que por si s� merece umas linhas. Outra raz�o � que eu admiro verdadeiramente as pessoas que escrevem este componente.... Eu fiz e ainda fa�o alguma programa��o, mas a minha experi�ncia anterior estava relacionada com aplica��es de neg�cio e actualmente � essencialmente scripts, pequenos utilit�rios, stored procedures etc. N�o querendo ofender ningu�m, mas honestamente acredito que estes dois mundos de programa��o est�o em campeonatos completamente diferentes. Tudo isto para dizer que o optimizador � algo admir�vel dada a complexidade dos problemas que ele tem de resolver.
Tamb�m tenho uma vis�o emp�rica ou apenas uma impress�o relativamente a alguns optimizadores no mercado de bases de dados. Esta impress�o deriva da minha experi�ncia pessoal, do meu contacto com profissionais que trabalham com outras tecnologias e de alguma leitura... Basicamente penso que o nosso maior competidor tem um optimizador com um comportamento insconstante, dif�cil de entender e pouco confi�vel. Uma das outras bases de dados relacionais da IBM, DB2 neste caso, tem provavelmente o optimizador mais "inteligente", mas n�o tenho practicamente nenhuma experi�ncia com ele. No caso do Informix, considero que n�o � o mais sofisticado (muitas melhorias na vers�o 10 e 11, e mais no Panther), mas realmente adoro a sua robustez. Tipicamente apenas precisa de estat�sticas. Julgo que s� utilizei ou recomendei optimizer directives duas vezes desde que trabalho com Informix. E que me lembre nunca bati pessoalmente num bug do optimizador (embora baste consultar a lista de correc��es de cada fixpack para sabermos que eles existem).
Para terminar esta introdu��o gostaria de referir que uma das minhas tarefas habituais � a optimiza��o de queries. Por isso j� encontrei algumas situa��es interessantes. Como escrevi acima, tipicamente um UPDATE STATISTICS resolve o problema. Algumas vezes encontrei limita��es do optimizador (as coisas poderiam correr melhor se ele pudesse resolver as queries de outra maneira) e at� limita��es do motor (o plano de execu��o gerado � bom, mas a execu��o leva tempo desnecess�rio a correr). A situa��o que se segue chegou-me � aten��o atrav�s de um cliente que estava a analisar queries sobre uma tabela que estava a sofrer muitos sequential scans.... Vejamos a defini��o da tabela primeiro:
create table example_table
(
et_id serial not null ,
et_ts datetime year to second,
et_col1 smallint,
et_col2 integer,
et_var1 varchar(50),
et_var2 varchar(50),
et_var3 varchar(50),
et_var4 varchar(50),
et_var5 varchar(50),
primary key (et_id)
)
lock mode row;
create index ids_example_table_1 on example_table (et_timestamp) using btree in dbs1
E agora dois planos de execu��o que obtiveram:
QUERY: (OPTIMIZATION TIMESTAMP: 08-25-2010 12:31:56)
------
SELECT et_ts, et_col2
FROM example_table WHERE et_var2='ValorMenosComum'
AND et_col1 = 1 ORDER BY et_ts DESC
Estimated Cost: 5439
Estimated # of Rows Returned: 684
Temporary Files Required For: Order By
1) informix.example_table: SEQUENTIAL SCAN
Filters: (informix.example_table.et_var2 = 'ValorMenosComum' AND informix.example_table.et_col1 = 1 )
Esta era uma das queries respons�vel pelos sequential scans. Nada de especial. As condi��es da query incidem em colunas n�o indexadas. Portanto n�o utiliza nenhum ind�ce. Inclui uma cl�usula ORDER BY, por isso um necessita de "ficheiros" tempor�rios para ordena��o. (na verdade pode ser feito em mem�ria, mas o plano quer dizer que precisa de fazer uma opera��o de ordena��o).
Portanto nada a salientar aqui.... Mas vejamos o outro plano sobre a mesma tabela. Na verdade a query � muito semelhante:
QUERY: (OPTIMIZATION TIMESTAMP: 08-25-2010 12:28:32)
------
SELECT et_ts, et_col2
FROM example_table WHERE et_var2='ValorMaisComum'
AND et_col1 = 1 ORDER BY et_ts DESC
Estimated Cost: 5915
Estimated # of Rows Returned: 2384
1) informix.example_table: INDEX PATH
Filters: (informix.example_table.et_var2 = 'ValorMaisComum' AND informix.example_table.et_col1 = 1 )
(1) Index Name: informix.ids_example_table_1
Index Keys: et_ts (Serial, fragments: ALL)
Ok... � quase a mesma query, mas desta vez temos um INDEX PATH... Um momento!.... N�o escrevi que as condi��es n�o indiciam sobre colunas indexadas? Sim. Outra diferen�a: Neste caso n�o necessita de "temporary files" (o que significa que n�o est� a fazer ordena��o). Note-se ainda que n�o existe nenhum filtro aplicado ao ind�ce. Portanto, neste caso est� a fazer algo completamente diferente da situa��o anterior. L� o ind�ce (todo o ind�ce), acede �s linhas (assim obt�m-nas j� ordenadas) e depois aplica os filtros descartando as linhas que n�o verificam as condi��es. Faz isto porque pretende evitar a opera��o de ordena��o. Mas porque � que muda o comportamento? Poder� ter reparado que j� inclui uma pista sobre isto. Para proteger a privacidade dos dados mudei o nome da tabela, dos campos e dos valores das queries. E no primeiro caso utilizei "ValorMenosComum" e no segundo usei "ValorMaisComum". Portanto como se pode esperar, o valor da segunda query � mais frequente na tabela que o valor usado na primeira query. Como sabemos isso? Primeiro olhando para o valor estimado de n�mero de linhas retornadas pela query (684 vs 2384). Como � que o optimizador prev� estes valores? Atrav�s das estat�sticas da tabela, em particular as distribui��es de cada coluna. Aqui est�o as mesmas para a coluna et_var2 (obtidas com dbschema -hd):
Distribution for informix.example_table.et_var2
Constructed on 2010-08-25 12:19:04.31216
Medium Mode, 2.500000 Resolution, 0.950000 Confidence
--- DISTRIBUTION ---
( OutroValor )
--- OVERFLOW ---
1: ( 4347, Value_1 )
2: ( 4231, Value_2 )
3: ( 3129, ValorMaisComum )
4: ( 2840, Value_4 )
5: ( 2405, Value_5 )
6: ( 3854, Value_6 )
7: ( 3941, Value_7 )
8: ( 2086, Value_8 )
9: ( 4086, Value_9 )
10: ( 4144, Value_10 )
11: ( 4086, Value_11 )
12: ( 2666, Value_12 )
13: ( 2869, Value_13 )
14: ( 3245, Value_14 )
15: ( 2811, Value_15 )
16: ( 3419, Value_16 )
17: ( 2637, Value_17 )
18: ( 3187, Value_18 )
19: ( 4347, Value_19 )
20: ( 898, ValorMenosComum )
21: ( 4144, Value_20 )
22: ( 2260, Value_21 )
23: ( 3999, Value_22 )
24: ( 1797, Value_23 )
25: ( 2115, Value_24 )
26: ( 2173, Value_25 )
27: ( 4144, Value_26 )
Para o caso de n�o estar habituado a analisar esta informa��o vou fazer uma pequena descri��o da mesma. As distribui��es s�o mostradas como uma lista de "cestos". Cada cesto, excepto o primeiro, cont�m 3 colunas: O n�mero de registo que representa, o n�mero de valores �nicos nesse cesto e o valor mais "alto" nesse cesto. O que est� acima n�o � um bom exemplo porque a sec��o de "cestos" normal s� tem um valor (OutroValor), e dado que � o primeiro apenas tem o valor, o que significa que � o valor mais "baixo" da coluna. Vejamos um exemplo retirado da descri�ao do utilit�rio dbschema no manual de migra��o:
( 5)
1: ( 16, 7, 11)
2: ( 16, 6, 17)
3: ( 16, 8, 25)
4: ( 16, 8, 38)
5: ( 16, 7, 52)
6: ( 16, 8, 73)
7: ( 16, 12, 95)
8: ( 16, 12, 139)
9: ( 16, 11, 182)
10: ( 10, 5, 200)
Portanto olhando para isto podemos ver que o valor mais baixo � 5, que entre 5 e 11 temos 16 valores, 7 dos quais s�o �nicos. Entre 11 e 17 temos 6 valores �nicos etc.
Voltando � nossa situa��o concreta vemos uma sec��o de "OVERFLOW". Esta sec��o � criada quando h� valores altamente repetidos que iriam deturpar ou desequilibrar as distribui��es. Nestes casos os valores repetidos s�o mostrados lado a lado com o n�mero de vezes que aparecem.
Assim, no nosso caso, estamos interessados nos valores "ValorMenosComum" (898 vezes) e "ValorMaisComum" (3129 vezes). Portanto, com estas evid�ncias podemos perceber a decis�o de escolher o INDEX PATH. Quando o motor chega ao final do conjunto de resultados o mesmo j� est� ordenado. E nesta situa��o (mais de 3000 linhas) a ordena��o consumiria mais recursos qeu para a outra situa��o em an�lise (cerca de 800 linhas).
Pode ser muito discutivel se a decis�o � efectivamente correcta. Mas o que quero salientar � que o optimizador se preocupa.... Ou seja, a an�lise que faz das condi��es e op��es � profunda e � capaz de escolher um plano de execu��o que n�o � �bvio (e escolh�-lo por boas raz�es).
