RE: row filtering for logical replication - Mailing list pgsql-hackers

On Saturday, January 29, 2022 8:31 AM Andres Freund <andres@anarazel.de> wrote:
> 
> Hi,
> 
> Are there any recent performance evaluations of the overhead of row filters? I
> think it'd be good to get some numbers comparing:

Thanks for looking at the patch! Will test it.

> 1) $workload with master
> 2) $workload with patch, but no row filters
> 3) $workload with patch, row filter matching everything
> 4) $workload with patch, row filter matching few rows
> 
> For workload I think it'd be worth testing:
> a) bulk COPY/INSERT into one table
> b) Many transactions doing small modifications to one table
> c) Many transactions targetting many different tables
> d) Interspersed DDL + small changes to a table
> > +/*
> > + * Initialize for row filter expression execution.
> > + */
> > +static ExprState *
> > +pgoutput_row_filter_init_expr(Node *rfnode) {
> > +    ExprState  *exprstate;
> > +    Expr       *expr;
> > +
> > +    /*
> > +     * This is the same code as ExecPrepareExpr() but that is not used
> because
> > +     * we want to cache the expression. There should probably be another
> > +     * function in the executor to handle the execution outside a normal
> Plan
> > +     * tree context.
> > +     */
> > +    expr = expression_planner((Expr *) rfnode);
> > +    exprstate = ExecInitExpr(expr, NULL);
> > +
> > +    return exprstate;
> > +}
> 
> In what memory context does this run? Are we taking care to deal with leaks?
> I'm pretty sure the planner relies on cleanup via memory contexts.

It was running under entry->cache_expr_cxt.

> > +    memset(entry->exprstate, 0, sizeof(entry->exprstate));
> > +
> > +    schemaId = get_rel_namespace(entry->publish_as_relid);
> > +    schemaPubids = GetSchemaPublications(schemaId);
> 
> Isn't this stuff that we've already queried before? If we re-fetch a lot of
> information it's not clear to me that it's actually a good idea to defer building
> the row filter.
> 
> 
> > +    am_partition = get_rel_relispartition(entry->publish_as_relid);
> 
> All this stuff likely can cause some memory "leakage" if you run it in a long-lived
> memory context.
> 
> 
> > +    /*
> > +     * Find if there are any row filters for this relation. If there are,
> > +     * then prepare the necessary ExprState and cache it in
> > +     * entry->exprstate. To build an expression state, we need to ensure
> > +     * the following:
...
> > +     *
> > +     * ALL TABLES IN SCHEMA implies "don't use row filter expression" if
> > +     * the schema is the same as the table schema.
> > +     */
> > +    foreach(lc, data->publications)
...
> > +        else if (list_member_oid(schemaPubids, pub->oid))
> > +        {
> > +            /*
> > +             * If the publication is FOR ALL TABLES IN SCHEMA and
> it overlaps
> > +             * with the current relation in the same schema then this
> is also
> > +             * treated same as if this table has no row filters (even if
> for
> > +             * other publications it does).
> > +             */
> > +            pub_no_filter = true;
> 
> Isn't this basically O(schemas * publications)?

Moved the row filter initialization code to get_rel_sync_entry.

> 
> > +    if (has_filter)
> > +    {
> > +        /* Create or reset the memory context for row filters */
> > +        if (entry->cache_expr_cxt == NULL)
> > +            entry->cache_expr_cxt =
> AllocSetContextCreate(CacheMemoryContext,
> > +
>                   "Row filter expressions",
> > +
>                   ALLOCSET_DEFAULT_SIZES);
> > +        else
> > +            MemoryContextReset(entry->cache_expr_cxt);
> 
> I see this started before this patch, but I don't think it's a great idea that
> pgoutput does a bunch of stuff in CacheMemoryContext. That makes it
> unnecessarily hard to debug leaks.
> 
> Seems like all this should live somwhere below ctx->context, allocated in
> pgoutput_startup()?
> 
> Consider what happens in a long-lived replication connection, where
> occasionally there's a transient error causing streaming to stop. At that point
> you'll just loose all knowledge of entry->cache_expr_cxt, no?
> 
> 
> > +
> > +/* Inialitize the slot for storing new and old tuple */ static void
> > +init_tuple_slot(Relation relation, RelationSyncEntry *entry) {
> > +    MemoryContext    oldctx;
> > +    TupleDesc        oldtupdesc;
> > +    TupleDesc        newtupdesc;
> > +
> > +    oldctx = MemoryContextSwitchTo(CacheMemoryContext);
> > +
> > +    /*
> > +     * Create tuple table slots. Create a copy of the TupleDesc as it needs to
> > +     * live as long as the cache remains.
> > +     */
> > +    oldtupdesc = CreateTupleDescCopy(RelationGetDescr(relation));
> > +    newtupdesc = CreateTupleDescCopy(RelationGetDescr(relation));
> > +
> > +    entry->old_slot = MakeSingleTupleTableSlot(oldtupdesc,
> &TTSOpsHeapTuple);
> > +    entry->new_slot = MakeSingleTupleTableSlot(newtupdesc,
> > +&TTSOpsHeapTuple);
> > +
> > +    MemoryContextSwitchTo(oldctx);
> > +}
> 
> This *definitely* shouldn't be allocated in CacheMemoryContext. It's one thing
> to have a named context below CacheMemoryContext, that's still somewhat
> identifiable. But allocating directly in CacheMemoryContext is almost always a
> bad idea.
> 
> What is supposed to clean any of this up in case of error?
> 
> 
> I guess I'll start a separate thread about memory handling in pgoutput :/

Thanks for the comments.
Added a separate memory context below ctx->context and
allocate all these newly added stuff under the separate memory context for now.

It seems you mean the existing stuff should also be put into a separate memory
context like this, do you think we can do it as a spearate patch or include
that change in row filter patch ?

> > +    /*
> > +     * We need this map to avoid relying on ReorderBufferChangeType
> enums
> > +     * having specific values.
> > +     */
> > +    static int map_changetype_pubaction[] = {
> > +        [REORDER_BUFFER_CHANGE_INSERT] = PUBACTION_INSERT,
> > +        [REORDER_BUFFER_CHANGE_UPDATE] = PUBACTION_UPDATE,
> > +        [REORDER_BUFFER_CHANGE_DELETE] = PUBACTION_DELETE
> > +    };
> 
> Why is this "static"? Function-local statics only really make sense for variables
> that are changed and should survive between calls to a function.

Removed the "static" label.

> > +    Assert(*action == REORDER_BUFFER_CHANGE_INSERT ||
> > +           *action == REORDER_BUFFER_CHANGE_UPDATE ||
> > +           *action == REORDER_BUFFER_CHANGE_DELETE);
> > +
> > +    Assert(new_slot || old_slot);
> > +
> > +    /* Get the corresponding row filter */
> > +    filter_exprstate =
> > +entry->exprstate[map_changetype_pubaction[*action]];
> > +
> > +    /* Bail out if there is no row filter */
> > +    if (!filter_exprstate)
> > +        return true;
> > +
> > +    elog(DEBUG3, "table \"%s.%s\" has row filter",
> > +         get_namespace_name(RelationGetNamespace(relation)),
> > +         RelationGetRelationName(relation));
> > +
> > +    estate = create_estate_for_relation(relation);
> > +    ecxt = GetPerTupleExprContext(estate);
> 
> So we do this for each filtered row? That's a *lot* of overhead.
> CreateExecutorState() creates its own memory context, allocates an EState,
> then GetPerTupleExprContext() allocates an ExprContext, which then creates
> another memory context.

Cached the estate in the new version.

> I don't really see any need to allocate this over-and-over?
> 
> >          case REORDER_BUFFER_CHANGE_INSERT:
> >              {
> > -                HeapTuple    tuple =
> &change->data.tp.newtuple->tuple;
> > +                /*
> > +                 * Schema should be sent before the logic that
> replaces the
> > +                 * relation because it also sends the ancestor's
> relation.
> > +                 */
> > +                maybe_send_schema(ctx, change, relation,
> relentry);
> > +
> > +                new_slot = relentry->new_slot;
> > +
> > +                ExecClearTuple(new_slot);
> > +
>     ExecStoreHeapTuple(&change->data.tp.newtuple->tuple,
> > +                                   new_slot,
> false);
> 
> Why? This isn't free, and you're doing it unconditionally. I'd bet this alone is
> noticeable slowdown over the current state.

It was intended to avoid deform the tuple twice, once in row filter execution ,second time
in logicalrep_write_tuple. But I will test the performance impact of this and improve
this if needed.

Best regards,
Hou zj