Обсуждение: Stream consistent snapshot via a logical decoding plugin as a series of INSERTs

I'd like to propose generic functions (probably in an extension, or in core if not possible otherwise) to facilitate streaming existing data from the database *in the same format* that one would get if these would be the changes decoded by a logical decoding plugin.

The idea is to use a snapshot returned from CREATE_REPLICATION_SLOT command of the replication protocol to get a consistent snapshot of the database, then start listening to new changes on the slot.

On 15 January 2016 at 08:30, Shulgin, Oleksandr <oleksandr.shulgin@zalando.de> wrote:

 

I'd like to propose generic functions (probably in an extension, or in core if not possible otherwise) to facilitate streaming existing data from the database *in the same format* that one would get if these would be the changes decoded by a logical decoding plugin.

The idea is to use a snapshot returned from CREATE_REPLICATION_SLOT command of the replication protocol to get a consistent snapshot of the database, then start listening to new changes on the slot.

It sounds like this is already possible. 

On Fri, Jan 15, 2016 at 11:08 AM, Simon Riggs <simon@2ndquadrant.com> wrote:

On 15 January 2016 at 08:30, Shulgin, Oleksandr <oleksandr.shulgin@zalando.de> wrote:

 

I'd like to propose generic functions (probably in an extension, or in core if not possible otherwise) to facilitate streaming existing data from the database *in the same format* that one would get if these would be the changes decoded by a logical decoding plugin.

The idea is to use a snapshot returned from CREATE_REPLICATION_SLOT command of the replication protocol to get a consistent snapshot of the database, then start listening to new changes on the slot.

It sounds like this is already possible. 

Totally, that's how it was supposed to be used anyway.  What is missing IMO is retrieving the initial snapshot in the same format as that the later changes will arrive.

POC patch attached.  Findings:

1) Needs an actual slot for all the decode machinery to work (code depends on MyReplicationSlot being set).

2) Requires a core patch.

3) Currently only supports textual output, adding binary is trivial.

Acquiring a slot means this cannot be run in parallel from multiple backends.  Any ideas on how to overcome this (except for opening multiple slots with the same LSN)?

To obtain a consistent snapshot, the client still needs to take care of preserving and setting transaction snapshot properly.

I'd like to propose generic functions (probably in an extension, or in core if not possible otherwise) to facilitate streaming existing data from the database *in the same format* that one would get if these would be the changes decoded by a logical decoding plugin.

The idea is to use a snapshot returned from CREATE_REPLICATION_SLOT command of the replication protocol to get a consistent snapshot of the database, then start listening to new changes on the slot.

The way this initial export phase is implemented there is by providing a SQL-callable set returning function which is using SPI to run "SELECT * FROM mytable" behind the scenes and runs the resulting tuples through the INSERT callback of the logical decoding plugin, which lives in the same loadable module as this SQL function.

Bottled Water logical decoding plugin uses binary protocol based on Avro data serialization library.  As an experiment I was adding support for JSON output format to it, and for that I had to re-implement the aforementioned SRF to export initial data to convert tuples to JSON instead.

What do you say?

On 15 January 2016 at 16:30, Shulgin, Oleksandr <oleksandr.shulgin@zalando.de> wrote:

 

I'd like to propose generic functions (probably in an extension, or in core if not possible otherwise) to facilitate streaming existing data from the database *in the same format* that one would get if these would be the changes decoded by a logical decoding plugin.

So effectively produce synthetic logical decoding callbacks to run a bunch of fake INSERTs, presumably with a fake xid etc?

The idea is to use a snapshot returned from CREATE_REPLICATION_SLOT command of the replication protocol to get a consistent snapshot of the database, then start listening to new changes on the slot.

My impression is that you want to avoid the current step of "synchronize database initial contents" when using logical decoding for replication.

But I guess you're looking to then populate that empty schema in-band via logical decoding, rather than having to do a --data-only dump or use COPY.  Right?

That won't help you for schema; presumably you'd still do a pg_dump --schema-only | pg_restore for that.

Just like when restoring a --data-only dump or using COPY you'd have to disable FKs during sync, but that's pretty much unavoidable.

The way this initial export phase is implemented there is by providing a SQL-callable set returning function which is using SPI to run "SELECT * FROM mytable" behind the scenes and runs the resulting tuples through the INSERT callback of the logical decoding plugin, which lives in the same loadable module as this SQL function.

o_O

What about the reorder buffer, the logical decoding memory context, etc?

Bottled Water logical decoding plugin uses binary protocol based on Avro data serialization library.  As an experiment I was adding support for JSON output format to it, and for that I had to re-implement the aforementioned SRF to export initial data to convert tuples to JSON instead.

Have you taken a look at what's been done with pglogical and pglogical_output? 

We've got extensible protocol support there, and if Avro offers compelling benefits over the current binary serialization I'm certainly interested in hearing about it.

What do you say?

Interesting idea. As outlined I think it sounds pretty fragile though; I really, really don't like the idea of lying to the insert callback by passing it a fake insert with (presumably) fake reorder buffer txn, etc.

What we've done in pglogical is take a --schema-only dump then, on the downstream, attach to the exported snapshot and use COPY ... TO STDOUT over a libpq connection to the upstream feed that to COPY ... FROM STDIN on another libpq connection to "ourselves", i.e. the downstream. Unless Petr changed it to use COPY innards directly on the downstream; I know he talked about it but haven't checked if he did. Anyway, either way it's not pretty and requires a sideband non-replication connection to sync initial state. The upside is that it can be relatively easily parallelized for faster sync using multiple connections.

To what extent are you setting up a true logical decoding context here?

Where does the xact info come from? The commit record? etc.

You're presumably not forming a reorder buffer then decoding it since it could create a massive tempfile on disk, so are you just dummying this info up?

Or hoping the plugin won't look at it?

The functionality is good and I think that for the SQL level you'd have to use SET TRANSACTION SNAPSHOT as you show. But I think it should really be usable from the replication protocol too - and should try to keep the state as close to that of a normal decoding session as possible. We'd at least need a new walsender protocol command equivalent that took the snapshot identifier, relation info and the other decoding params instead of a slot name. Or, ideally, a variant on START_REPLICATION ... LOGICAL ... that omits SLOT and instead takes TABLES as an argument, with a list of relation(s) to sync. Unlike normal START_REPLICATION ... LOGICAL ... it'd return to walsender protocol mode on completion, like the phys rep protocol does when it's time for a timeline switch.

Rather than lie to the insert callback I'd really rather define a new logical decoding callback for copying initial records. It doesn't get any xact info (since it's not useful/relevant) or a full reorder buffer. No ReorderBufferChange is passed; instead we pass something like a ReorderBufferSync that contains the new tuple ReorderBufferTupleBuf, origin id, origin lsn and commit timestamp (if known) and the RelFileNode affected. The LogicalDecodingContext that's set up for the callback gets ctx->reorder = NULL .  There's no ReorderBufferTxn argument and none is defined.

Since it's a new callback the plugin knows the rules, knows it's getting initial state data to sync over, etc. It doesn't have to try to guess if it's seeing a real insert and act differently with respect to xact identity etc.

Obviously that's 9.6 material at the soonest, and only 9.6 if it could be done ... well, right about now. So that won't meet your immediate needs, but I think the same is true of the interface you propose above.

What I suggest doing in the mean time is specifying a new callback function interface for tuple copies as described above, to be implemented by logical decoding modules that support this extension. In each decoding plugin we then define a SQL-callable function with 'internal' return type that returns a pointer to the callback so you can obtain the hook function address via a fmgr call via pg_proc. The callback would expect a state much like I describe above and we'd use a SQL-callable function like what you outlined to set up a fake logical decoding state for it, complete with decoding context etc. Probably copying & pasting a moderately painful amount of the logical decoding guts into an ext in the process :( since I don't think you can easily set up much of the decoding state using the decoding backend code without having a slot to use. Still, that'd let us prototype this and prove the idea for inclusion in 9.7 (?) in-core while retaining the capability via an extension for earlier versions.

You'd have to do much of the same hoop jumping to call an arbitrary output plugin's insert callback directly, if not more.

Alternately, you could just use COPY ;)

All of this implies another *postgres* database on the receiving side, which is not necessarily the case for my research.

This is what I'm going to benchmark.  With the generic function I can just create two slots: one for pglogical and another one for BottledWater/Avro and see which one performs better when forced to stream some TB worth of INSERTs through the change callback.

What do you say?

Interesting idea. As outlined I think it sounds pretty fragile though; I really, really don't like the idea of lying to the insert callback by passing it a fake insert with (presumably) fake reorder buffer txn, etc.

Fair enough.  However for performance testing it could be not that bad, even if nothing of that lands in the actual API.

You're presumably not forming a reorder buffer then decoding it since it could create a massive tempfile on disk, so are you just dummying this info up?

In my experience, it doesn't.  We know it's going to be a "committed xact", so we don't really need to queue the changes up before we see a "commit" record.

 

Another consideration is that we might introduce modes for acquiring the slot: Exclusive and Shared access (can be implemented with LWLocks?), so that peek_changes() and stream_relation() could acquire the slot in Shared access mode, thus allowing parallel queries, while START_REPLICATION and get_changes() would require Exclusive access.

Thanks for the thoughtful reply!  I'm going to experiment with my toy code a bit more, while keeping in mind what could a more workable approach look like.

On 20 January 2016 at 15:50, Shulgin, Oleksandr <oleksandr.shulgin@zalando.de> wrote:

That'd be nice, but probably not totally necessary for streaming relations. It doesn't really need the slot at all. Or shouldn't, I think. Though it might be easiest to allow it to acquire the slot just for convenience and shared code.