Re: Online enabling of page level checksums

So as mentioned on IRC, I have a patch that I am working to rebase on HEAD with the following design. It is very similar to what you have proposed, so maybe we can use my development notes as a jumping-off point for discussion/refinement.

* Incremental Checksums

PostgreSQL users should have a way up upgrading their cluster to use data checksums without having to do a costly pg_dump/pg_restore; in particular, checksums should be able to be enabled/disabled at will, with the database enforcing the logic of whether the pages considered for a given database are valid.

Considered approaches for this are having additional flags to pg_upgrade to set up the new cluster to use checksums where they did not before (or optionally turning these off). This approach is a nice tool to have, but in order to be able to support this process in a manner which has the database online while the database is going throught the initial checksum process.

In order to support the idea of incremental checksums, this design adds the following things:

** pg_control:

Keep "data_checksum_version", but have it indicate *only* the algorithm version for checksums. i.e., it's no longer used for the data_checksum enabled/disabled state.

Add "data_checksum_state", an enum with multiple states: "disabled", "enabling", "enforcing" (perhaps "revalidating" too; something to indicate that we are reprocessing a database that purports to have been completely checksummed already)

An explanation of the states as well as the behavior of the checksums for each.

- disabled => not in a checksum cycle; no read validation, no checksums written. This is the current behavior for Postgres *without* checksums.

- enabling => in a checksum cycle; no read validation, write checksums. Any page that gets written to disk will be a valid checksum. This is required when transitioning a cluster which has never had checksums, as the page reads would normally fail since they are uninitialized.

- enforcing => not in a checksum cycle; read validation, write checksums. This is the current behavior of Postgres *with* checksums.

(caveat: I'm not certain the following state is needed (and the current version of this patch doesn't have it)):

- revalidating => in a checksum cycle; read validation, write checksums. The difference between this and "enabling" is that we care if page reads fail, since by definition they should have been validly checksummed, as we should verify this.

Add "data_checksum_cycle", a counter that gets incremented with every checksum cycle change. This is used as a flag to verify when new checksum actions take place, for instance if we wanted to upgrade/change the checksum algorithm, or if we just want to support periodic checksum validation.

This variable will be compared against new values in the system tables to keep track of which relations still need to be checksummed in the cluster.

** pg_database:

Add a field "datlastchecksum" which will be the last checksum cycle which has completed for all relations in that database.

** pg_class:

Add a field "rellastchecksum" which stores the last successful checksum cycle for each relation.

** The checksum bgworker:

When the enabling event is initiated, we will iterate over all databases, checking for all databases where the "datlastchecksum" field is < the current checksum cycle. For each of these, we will spawn a bgworker to connect to these dbs and iterate over pg_class looking for "rellastchecksum < data_checksum_cycle". If it finds none (i.e., every record has rellastchecksum == data_checksum_cycle) then it marks the containing database as up-to-date by updating "datlastchecksum = data_checksum_cycle". We can presumably skip over temporary and unlogged relations here.

For any relation that it finds in the database which is not checksummed, it starts an actual worker to handle the checksum process for this table. Since the state of the cluster is already either "enforcing" or "revalidating", any block writes will get checksums added automatically, so the only thing the bgworker needs to do is load each block in the relation and explicitly mark as dirty (unless that's not required for FlushBuffer() to do its thing). After every block in the relation is visited this way and checksummed, its pg_class record will have "rellastchecksum" updated.

(XXX: how to handle databases where connections are disabled, like "template1"?)

When all database have "datlastchecksum" == data_checksum_cycle, we initiate checksumming of any global cluster heap files. When the global cluster tables heap files have been checksummed, then we consider the checksum cycle complete, change pg_control's "data_checksum_state" to "enforcing" and consider things fully up-to-date.

** Function API:

Interface to the functionality will be via the following Utility functions:

- pg_enable_checksums(void) => turn checksums on for a cluster. Will error if the state is anything but "disabled". If this is the first time this cluster has run this, this will initialize ControlFile->data_checksum_version to the preferred built-in algorithm (since there's only one currently, we just set it to 1). This increments the ControlFile->data_checksum_cycle variable, then sets the state to "enabling", which means that the next time the bgworker checks if there is anything to do it will see that state, scan all the databases' "datlastchecksum" fields, and start kicking off the bgworker processes to handle the checksumming of the actual relation files.

- pg_disable_checksums(void) => turn checksums off for a cluster. Sets the state to "disabled", which means bg_worker will not do anything.

- pg_request_checksum_cycle(void) => if checksums are "enabled", increment the data_checksum_cycle counter and set the state to "enabling". (Alterantely, if we use the "revalidate" state here we could ensure that existing checksums are validated on read to alert us of any blocks with problems. This could also be made to be "smart" i.e., interrupt existing running checksum cycle to kick off another one (not sure of the use case), effectively call pg_enable_checksums() if the cluster has not been explictly enabled before, etc; depends on how pedantic we want to be.

** Design notes/implications:

When the system is in one of the modes which write checksums (currently everything but "disabled") any new relations/databases will have their "rellastchecksum"/"datlastchecksum" counters prepopulated with the current value of "data_checksum_counter", as we know that any space used for these relations will be checksummed, and hence valid. By pre-setting this, we remove the need for the checksum bgworker to explicitly visit these new relations and force checksums which will already be valid.

With checksums on, we know any full-heap-modifying operations will be properly checksummed, we may be able to pre-set rellastchecksum for other operations such as ALTER TABLEs which trigger a full rewrite *without* having to explicitly have the checksum bgworker run on this. I suspect there are a number of other places which may lend themselves to optimization like this to avoid having to process relations explicitly. (Say, if we somehow were able to force a checksum operation on any full SeqScan and update the state after the fact, we'd avoid paying this penalty another time.)

** pg_upgrade:

With this additional complexity, we need to consider pg_upgrade, both now and in future versions. For one thing, we need to transfer settings from pg_control, plus make sure that pg_upgrade accepts deviances in any of the data_checksum-related settings in pg_control.

4 scenarios to consider if/what to allow:

*** non-checksummed -> non-checksummed

exactly as it stands now

*** checksummed -> non-checksummed

pretty trivial; since the system tables will be non-checksummed, just equivalent to resetting the checksum_cycle and pg_control fields; user data files will be copied or linked into place with the checksums, but since it is disbled they will be ignored.

*** non-checksummed -> checksummed

For the major version this patch makes it into, this will likely be the primary use case; add an --enable-checksums option to `pg_upgrade` to initially set the new cluster state to the checksums_enabling state, pre-init the system databases with the correct state and checksum cycle flag, or have checksums be the *default* option. Either way, this needs to be set in the initial system initialization.

*** checksummed -> checksummed

The potentially tricky case (but likely to be more common going forward as incremental checksums are supported).

Since we may have had checksum cycles in process in the old cluster or otherwise had the checksum counter we need to do the following:

- need to propagate data_checksum_state, data_checksum_cycle, and data_checksum_version. If we wanted to support a different CRC algorithm, we could pre-set the data_checksum_version to a different version here, increment data_checksum_cycle, and set data_checksum_state to either "enabling" or "revalidating", depending on the original state from the old cluster. (i.e., if we were in the middle of an initial checksum cycle (state == "enabling").

- new cluster's system tables may need to have the "rellastchecksum" and "datlastchecksum" settings saved from the previous system, if that's easy, to avoid a fresh checksum run if there is no need.

** Handling checksums on a standby:

How to handle checksums on a standby is a bit trickier since checksums are inherently a local cluster state and not WAL logged but we are storing state in the system tables for each database we need to make sure that the replicas reflect truthful state for the checksums for the cluster.

In order to manage this discrepency, we WAL log a few additional pieces of information; specifically:

- new events to capture/propogate any of the pg_control fields, such as: checksum version data, checksum cycle increases, enabling/disabling actions

- checksum background worker block ranges. (XXX: we could decide that we're okay with relations being all or nothing here, rendering this point moot.)

Some notes on the block ranges: This would effectively be a series of records containing (datid, relid, start block, end block) for explicit checksum ranges, generated by the checksum bgworker as it checksums individual relations. This could be broken up into a series of blocks so rather than having the granularity be by relation we could have these records get generated periodicaly (say in groups of 10K blocks or whatever, number to be determined) to allow standby checksum recalculation to be incremental so as not to delay replay unnecessarily as checksums are being created.

Since the block range WAL records will be replayed before any of the pg_class/pg_database catalog records are replay, we'll be guaranteed to have the checksums calculated on the standby by the time it appears valid due to system state.

We may also be able to use the WAL records to speed up the processing of existing heap files if they are interrupted for some reason, this remains to be seen.

** Testing changes:

We need to add separate initdb checksum regression test which are outside of the normal pg_regress framework.

--
David Christensen
End Point Corporation
david(at)endpoint(dot)com
785-727-1171