mirror of https://github.com/postgres/postgres
parent
a4155d3bbd
commit
7809cbf7e6
@ -0,0 +1,99 @@ |
|||||||
|
$Header: /cvsroot/pgsql/src/backend/executor/README,v 1.1 2001/05/15 00:35:50 tgl Exp $ |
||||||
|
|
||||||
|
The Postgres Executor |
||||||
|
--------------------- |
||||||
|
|
||||||
|
The executor processes a tree of "plan nodes". The plan tree is essentially |
||||||
|
a demand-pull pipeline of tuple processing operations. Each node, when |
||||||
|
called, will produce the next tuple in its output sequence, or NULL if no |
||||||
|
more tuples are available. If the node is not a primitive relation-scanning |
||||||
|
node, it will have child node(s) that it calls in turn to obtain input |
||||||
|
tuples. |
||||||
|
|
||||||
|
Refinements on this basic model include: |
||||||
|
|
||||||
|
* Choice of scan direction (forwards or backwards). Caution: this is not |
||||||
|
currently well-supported. It works for primitive scan nodes, but not very |
||||||
|
well for joins, aggregates, etc. |
||||||
|
|
||||||
|
* Rescan command to reset a node and make it generate its output sequence |
||||||
|
over again. |
||||||
|
|
||||||
|
* Parameters that can alter a node's results. After adjusting a parameter, |
||||||
|
the rescan command must be applied to that node and all nodes above it. |
||||||
|
There is a moderately intelligent scheme to avoid rescanning nodes |
||||||
|
unnecessarily (for example, Sort does not rescan its input if no parameters |
||||||
|
of the input have changed, since it can just reread its stored sorted data). |
||||||
|
|
||||||
|
The plan tree concept implements SELECT directly: it is only necessary to |
||||||
|
deliver the top-level result tuples to the client, or insert them into |
||||||
|
another table in the case of INSERT ... SELECT. (INSERT ... VALUES is |
||||||
|
handled similarly, but the plan tree is just a Result node with no source |
||||||
|
tables.) For UPDATE, the plan tree selects the tuples that need to be |
||||||
|
updated (WHERE condition) and delivers a new calculated tuple value for each |
||||||
|
such tuple, plus a "junk" (hidden) tuple CTID identifying the target tuple. |
||||||
|
The executor's top level then uses this information to update the correct |
||||||
|
tuple. DELETE is similar to UPDATE except that only a CTID need be |
||||||
|
delivered by the plan tree. |
||||||
|
|
||||||
|
XXX a great deal more documentation needs to be written here... |
||||||
|
|
||||||
|
|
||||||
|
EvalPlanQual (READ COMMITTED update checking) |
||||||
|
--------------------------------------------- |
||||||
|
|
||||||
|
For simple SELECTs, the executor need only pay attention to tuples that are |
||||||
|
valid according to the snapshot seen by the current transaction (ie, they |
||||||
|
were inserted by a previously committed transaction, and not deleted by any |
||||||
|
previously committed transaction). However, for UPDATE and DELETE it is not |
||||||
|
cool to modify or delete a tuple that's been modified by an open or |
||||||
|
concurrently-committed transaction. If we are running in SERIALIZABLE |
||||||
|
isolation level then we just raise an error when this condition is seen to |
||||||
|
occur. In READ COMMITTED isolation level, we must work a lot harder. |
||||||
|
|
||||||
|
The basic idea in READ COMMITTED mode is to take the modified tuple |
||||||
|
committed by the concurrent transaction (after waiting for it to commit, |
||||||
|
if need be) and re-evaluate the query qualifications to see if it would |
||||||
|
still meet the quals. If so, we regenerate the updated tuple (if we are |
||||||
|
doing an UPDATE) from the modified tuple, and finally update/delete the |
||||||
|
modified tuple. SELECT FOR UPDATE behaves similarly, except that its action |
||||||
|
is just to mark the modified tuple for update by the current transaction. |
||||||
|
|
||||||
|
To implement this checking, we actually re-run the entire query from scratch |
||||||
|
for each modified tuple, but with the scan node that sourced the original |
||||||
|
tuple set to return only the modified tuple, not the original tuple or any |
||||||
|
of the rest of the relation. If this query returns a tuple, then the |
||||||
|
modified tuple passes the quals (and the query output is the suitably |
||||||
|
modified update tuple, if we're doing UPDATE). If no tuple is returned, |
||||||
|
then the modified tuple fails the quals, so we ignore it and continue the |
||||||
|
original query. (This is reasonably efficient for simple queries, but may |
||||||
|
be horribly slow for joins. A better design would be nice; one thought for |
||||||
|
future investigation is to treat the tuple substitution like a parameter, |
||||||
|
so that we can avoid rescanning unrelated nodes.) |
||||||
|
|
||||||
|
Note a fundamental bogosity of this approach: if the relation containing |
||||||
|
the original tuple is being used in a self-join, the other instance(s) of |
||||||
|
the relation will be treated as still containing the original tuple, whereas |
||||||
|
logical consistency would demand that the modified tuple appear in them too. |
||||||
|
But we'd have to actually substitute the modified tuple for the original, |
||||||
|
while still returning all the rest of the relation, to ensure consistent |
||||||
|
answers. Implementing this correctly is a task for future work. |
||||||
|
|
||||||
|
In UPDATE/DELETE, only the target relation needs to be handled this way, |
||||||
|
so only one special recheck query needs to execute at a time. In SELECT FOR |
||||||
|
UPDATE, there may be multiple relations flagged FOR UPDATE, so it's possible |
||||||
|
that while we are executing a recheck query for one modified tuple, we will |
||||||
|
hit another modified tuple in another relation. In this case we "stack up" |
||||||
|
recheck queries: a sub-recheck query is spawned in which both the first and |
||||||
|
second modified tuples will be returned as the only components of their |
||||||
|
relations. (In event of success, all these modified tuples will be marked |
||||||
|
for update.) Again, this isn't necessarily quite the right thing ... but in |
||||||
|
simple cases it works. Potentially, recheck queries could get nested to the |
||||||
|
depth of the number of FOR UPDATE relations in the query. |
||||||
|
|
||||||
|
It should be noted also that UPDATE/DELETE expect at most one tuple to |
||||||
|
result from the modified query, whereas in the FOR UPDATE case it's possible |
||||||
|
for multiple tuples to result (since we could be dealing with a join in |
||||||
|
which multiple tuples join to the modified tuple). We want FOR UPDATE to |
||||||
|
mark all relevant tuples, so we pass all tuples output by all the stacked |
||||||
|
recheck queries back to the executor toplevel for marking. |
Loading…
Reference in new issue