The reasoning in the FAQ and code about RFC-2818 is only valid when using host style URIs. According to AWS S3 bucket naming standard the allowed characters is any lowercase alphanumeric including dash and dot.
Most self hosted S3 services utilize path based URIs where dots are valid in a bucket name so this check should only apply for host based buckets. Even though its use is not recommended and path-based access is being phased out of AWS S3 it is still valid and should function for other providers.
Options with unresolved dependencies can have an implied default specified. This makes the code a bit simpler since we don't need to check for option validity.
However, there was an edge case where if an option was specified in the config file and ultimately the dependency was not resolved then the option would not be marked as default and therefore show up in the option logging at the beginning of a command. The default value was correct so everything operated as expected but the logging was confusing.
In the case of an implied default, reinitialize the option struct so that any leftover settings will be reset.
The prior code allocated the entire chunk buffer when the file was opened. However, in practice many files are smaller than the chunk buffer, especially in the main process.
Instead grow the chunk buffer as data comes in to save memory when smaller files are being processed. This adds some overhead for reallocations but modern processors do this very efficiently so it should not be significant compared to the cost of compressing, encrypting, and transferring files. Even so, the growth is fairly aggressive when the input buffers are full so only one or two reallocation are required to get to the default chunk size.
Previously an S3 upload with default repo-storage-upload-chunk-size would only work for files <= 50GiB because of the limited number of chunks allowed. GCS has a smaller chunk size default so it topped out at 40GiB. Azure allows 50,000 chunks so it allowed up to 200GiB.
These are all far larger than files PostgreSQL will create but these days a data directory might also contain files created by plugins that can be much larger.
Since the eventual file size is not known in advance (due to compression) it is hard to pick an appropriate chunk size in advance. Instead, dynamically grow the chunk size over time to reach 5TiB for S3 and GCS (their upper limit). Azure has more parts so it will reach 45TiB, which is smaller than the upper limit of 190TiB, but seems sufficient for now.
The default buffer size is used for the first GiB (plus some) to provide compatibility with any clones that do not support variable block sizes. There is no evidence that this is a problem but better to be safe.
The minimum values for repo-storage-upload-chunk-size have been increased to match vendor minimums and simply the chunk size algorithm.
Since PostgreSQL 10 these settings have been defaulted to values required by the user guide so there is no need to explicitly set them.
PostgreSQL 9.5/9.6 are still supported by pgBackRest but are not represented in the user guide since they are EOL.
32-bit testing was broken by 24802a08, which was attempting to fix multi-architecture builds by using docker to set the architecture.
i386 is not a special case but the prior alternate architectures did not run integration tests. This requires passing the architecture around since the integration test main process runs on the host system, which may be a different architecture.
This makes maintenance easier. Also fix the command list for db-timeout so it matches pg-database, i.e. all the commands that can connect to the database.
Add +inherit, +role, and -command to help with command maintenance. These allow command lists to automatically add new commands without them needing to be added manually. They should also be easier to read than long command lists.
In many cases the valid commands are based on the commands valid for roles. In these cases derive the commands from a role list rather than an explicit command list.
Not only is this notation more compact but it helps prevent new commands from being missed.
This exposed a few issues:
1) The cmd option should only be valid when a command supports the local role since it is used to execute the local process. A number of commands were included before that did not have the local role.
2) cmd-ssh should be valid for any command that allows remotes. The annotate command was missing from this list.
3) compress-level-network should be valid for any command that allows remotes. The repo-rm command was missing from this list.
Restoring to a remote pg-host is not supported but the options were a bit untidy. Many options were marked as internal but should be invalid. repo-host-type and repo-host are required to let restore know if a pg-host is configured and remain internal but the rest of the pg-host-* options are now invalid for restore.
The same applies to the archive-get and archive-push commands although these were less likely to cause confusion.
Also reverse the dependency of pg-host and pg-host-type, i.e. make pg-host-type depend on pg-host, and alter pg-host-cmd and pg-host-user to depend on pg-host-type=ssh.
931435c0 added the ability to backup to a remote repo but did not quite get the option updates right. It worked, but a number of options were marked as internal so would not be visible to the user in command-line help.
Also reverse the dependency of repo-host and repo-host-type, i.e. make repo-host-type depend on repo-host, and alter repo-host-cmd and repo-host-user to depend on repo-host-type=ssh.
Previously internal state was not included because it does not affect how commands/options are used -- only whether they are shown in help. However, this makes it hard to know when the internal state changes because help is generated at build time and in any case is just a binary blob.
Internal state is not stored in the config structures since the macros resolve to nothing but it is handy for debugging to see when internal state has changed.
The Perl processing of config.yaml put hard limits on the format of that file. To allow flexibility in the file format remove all Perl processing on config.yaml.
This is just the beginning of migrating the preprocessor to C but even this small bit allows the removal of a lot of Perl code.
55e9969 updated the meson version but only reverted one of the changes implemented in 0eccbc8 where the version was lowered to >= 0.45.
Use get_option() as allowed by the updated version for clearer and more robust build code.
Per our policy to support five EOL versions of PostgreSQL, 9.5 is no longer supported by pgBackRest. Remove all logic associated with 9.5 and update the tests.
An effort was made to advance versions as much as possible in the tests while still providing coverage. Hopefully this will reduce churn when future versions expire, though it has created a bit more here.
Tests for 9.4/9.5 are left in the expire/info tests to demonstrate that these commands work with old versions present.
The 9.6 pg_control struct was being used for 9.5. This was not detected by testing because the new field introduced for 9.6 fit into an alignment hole in the 9.5 struct so the size of the struct and offset of all other members did not change. Since the new member was not used there was no impact on functionality.
9.5 is being removed in the next release so the only reason to fix this is to make the diff for that change more sensible, and to document that this happened.
Some of these types were versioned at one time. Others were not but it seemed better to version all of them for consistency. In fact this just creates churn when PostgreSQL versions are expired.
Also move the uint64 type to version.vendor.h since it is only used by versioned types.
Some of these macros have required different implementations according to PostgreSQL version but none of them do now.
The idea of versioning them even when not required was that it would reduce churn but in fact churn is increased because PostgreSQL versions expire far more often than these macros need version specific implementations.
Bug Fixes:
* Unnest HTTP/TLS/socket timeouts. (Reviewed by David Christensen.)
* Fix possible segfault in page checksum error message. (Fixed by Zsolt Parragi. Reviewed by David Steele.)
Features:
* Add repo-symlink option to suppress creation of repository symlinks. (Reviewed by Douglas J Hunley. Suggested by Ron Johnson.)
Improvements:
* Add HTTP retries for 408 and 429 errors. (Reviewed by David Christensen.)
On recent versions of Docker it is not necessary to specify the architecture of the require image. The --platform option takes care of it.
Rebuild test images that were modified by this change.
This test was removed in c64cd8e0 because it was taking too long to be scheduled, which held up testing.
Now that Github Actions supports aarch64 move the test there.
Similar to adc5e5b23, STRDEF assignment to reason can be out of scope, which may lead to either garbage in the log message or a crash with SIGSEGV.
Fix by using simple char * constants for the reason.
The SFTP storage driver did not set the pathSync method but had logic to see if pathSync was set. This led to some tortured tests since pathSync had to be injected into the interface after the storage object had been created.
Remove the pathSync parameter from storageWriteSftpNew() and also remove related tests.
These timeouts were nested, i.e. an error in the socket connection would also retry in the TLS and HTTP layers. This led to a multiplying effect such that it took nine minutes to fully timeout with the default io-timeout of 60 seconds.
A fix for this was attempted in 5314dbff but it reduced retries by too much and had to be reverted in fa5b2d44.
Instead fix by moving connection attempts to the lower layer (e.g. TLS -> socket) out of the exception block but leave it within the retry loop. So, for example, if the socket connection fails after retries then the error will not be retried by TLS. But if the TLS session fails the socket will be reconnected.
The STRDEF definition of plural is out of scope in the error message just below it, which may lead to either garbage in the log message or a crash with SIGSEGV.
This is particularly visible with pg_tde, which encrypts tables: if pgbackrest is run with checksum enabled, it tries to emit this message for all encrypted files.
Fix by using simple char * constants in the error message.
Rebuild all containers to get the most recent versions of PostgreSQL.
Update the Debian repository install to match current recommendations. This has already been done for the documentation in fcd00a45.
HTTP client errors 408 and 429 were not being retried but there seems to be some benefit to doing so.
408 has only been seen once and in that case the server was returning 400 (with Request Timeout in the response body) but it seems worth doing since this could happen during times of high congestion. Requests are not sent until request content is ready so there is not much else to be done to handle this error.
429 has been seen occasionally in the past but now seems to be common on Cloudflare R2. Ideally we would have a large initial back off here but it is not clear that is worth it at this time. The existing Fibonacci back off should be enough to allow operations to proceed if possible during the configured timeout.
Joakim Hindersson
David Steele
Stefan Fercot
Zsolt Parragi