Coverity complained that this pass by value was inefficient:
CID 376402: Performance inefficiencies (PASS_BY_VALUE)
Passing parameter file of type "ManifestFile" (size 136 bytes) by value.
This was completely intentional since it gives us a copy of the struct that we can change without bothering the caller. However, updating fields is fine and may benefit the caller at some future data, and in any case does no harm now.
And as usual it is easier not to fight with Coverity.
As much as possible it is better to get coverage with more realistic tests. Merging these modules will allow the page checksum code to be covered with real backups.
Limit which files can be added to bundles, which allows resume to work reasonably well. On resume, the bundles are removed and any remaining file is eligible to be to be resumed.
Also reduce the bundle-size default to 20MiB. This is pretty arbitrary, but a smaller default seems better.
Bundle (combine) smaller files during backup to reduce the number of files written to the repository (enable with --bundle). Reducing the number of files is a benefit on all file systems, but especially so on object stores such as S3 that have a high file creation cost. Another benefit is that zero-length files are only stored as metadata in the manifest.
Files are batched up to bundle-size and then compressed/encrypted individually and stored sequentially in the bundle. The bundle id and offset of each file is stored in the manifest so files can be retrieved randomly without needing to read the entire bundle. Files are ordered by timestamp descending when being assigned to bundles to reduce the amount of random access that needs to be done. The idea is that bundles with older files can be read in their entirety on restore and only bundles with newer files will get fragmented.
Bundles are a custom format with metadata stored in the manifest. Tar was considered but it is too limited a format, the major issue being that the size of the file must be known in advance and that is very contrary to how pgBackRest works, especially once we introduce page-level incremental backups.
Bundles are stored numbered in the bundle directory. Some files may still end up in pg_data if they are added after the backup is complete. backup_label is an example.
Currently, only the backup command works in batches. The restore and verify commands use the offsets to pull individual files out of the bundle. It seems better to finalize how this is going to work before optimizing the other commands. Even as is, this is a major step forward, and all commands function with bundling.
One caveat: resume is currently not supported when bundle is enabled.
There is some evidence that retrying fatal errors, especially out of memory errors, may cause lockups. It makes sense to report fatal errors as quickly as possible and bypass retries. This may or not fix the lockup issue but it is worth doing either way.
For now, the only fatal errors will be AssertError and MemoryError.
If the entire batch failed it would be retried, but individual file errors were not retried. This could cause pgBackRest to terminate during expiration or when removing an unresumable backup.
Rather than retry the entire batch, delete the errored files individually to take advantage of the HTTP retry rather than adding a new retry loop. These errors seem rare enough that it should not be a performance issue.
In theory, the additional stat() call after a file has been copied to the repo can determine if additional compression has been applied by the file system. However, it has been a very long time since we tested this in practice. There are currently no unit tests that accurately test this feature since it requires a compressed file system like ZFS to work, which never seemed worth the extra cost.
It can also add a lot of time to backups if there are a large quantity of small files.
In addition, it stands as a blocker for combining files for small file support since it is no longer possible to get per-file sizes from the viewpoint of the file system. There are several ways this could be reworked but none of them are easy while at the same time maintaining current info functionality.
It doesn't seem worth keeping an untested feature that will only work in some special cases (if it still works) when it is blocking development.
Coverity pointed out that a negative number could be passed to close(), which means the lock file would not get closed until the process ended. Proper execution does not require the file to be closed, but it is better to correctly free resources that are no longer needed.
The most recent release of Minio has broken CI builds but there is no logging to indicate what is wrong.
For now, just use the prior release to get CI builds working again. This kind if breakage is not uncommon for Minio but they usually resolve it in the next release.
Update lock code to use standard common/io functions and module patterns. This module was developed before the common/io module existed and our patterns had stabilized.
The /etc/profile.d/lang.sh script was causing issues but it does not exist on amd64, so it seems the easiest thing was to remove it.
Fix how 32-bit VMs are determined now that another 64-bit architecture has been added.
And remove some obsolete VM hashes.
Previously manifest load required two passes through the file list, one to load the data and one to set the defaults. This required each file to be packed twice.
Instead simply note that the file value is default and then set the file defaults when they are loaded from the manifest. This is made possible by the different internal/external representations for files so the same method cannot be applied to paths and links.
This change seems to resolve the performance issues noted in 61ce586 but there is no obvious reason why.
Centralize these options so they are consistent across clusters.
Also, there were some options that the user doesn't really need to see, .e.g. log_line_prefix. These can be set in advance so they don't need to be part of the documentation.
This function (which creates lots of tables) is generally useful for testing (not just stress testing) so create it as soon as the cluster is created.
Also add the data parameter which will insert a single row into the table so the file on disk is not zero bytes.
Manifests with a very large number of files can use a considerable amount of memory. There are a lot of zeroes in the data so it can be stored more efficiently by using base-128 varint encoding for the integers and storing the strings in the same allocation.
The downside is that the data needs to be unpacked in order to be used, but in most cases this seems fast enough (about 10% slower than before) except for saving the manifest, which is 10% slower up to 10 million files and then gets about 5x slower by 100 million (two minutes on my M1 Mac). Profiling does not show this slowdown so I wonder if this is related to the change in memory layout. Curiously, the function that increased most was jsonFromStrInternal(), which was not modified. That gives more weight to the idea that there is some kind of memory issue going on here and one hopes that servers would be less affected. Either way, they largest use cases we have seen are for about 6 million files so if we can improve that case I believe we will be better off.
Further analysis showed that most of the time was taken up writing the size and timestamp fields, which makes almost no sense. The same amount of time was used if they were hard-coded to 0, which points to some odd memory issue on the M1 architecture.
This change has been planned for a while, but the particular impetus at this time is that small file support requires additional fields that would increase manifest memory usage by about 20%, even if the feature is not used.
Note that the Pack code has been updated to use the new varint encoder, but the decoder remains separate because it needs to fetch one byte at a time.
Manifest defaults for user, group, and mode were previously generated by scanning the data to find the most common values. This was very accurate but slow and complicated. It could also lead to surprising changes in the manifest when a default value suddenly changed.
Instead, use the $PGDATA path to generate defaults. In the vast majority of cases the same user/group should own all the path/files and the default file mode is easily derived from the path mode. There may be some edge cases where this generates larger manifests, but in general it reduces time and complexity when saving the manifest.
Remove the MCV code since it is longer longer used.
Change the mode back to 0700 earlier to reduce churn in the expect logs.
This will be especially important in a future commit that gets the defaults exclusively from the base path.
This flag was only being used by the backup command after manifestNewBuild() and had no other uses. There was a time when it was important for integration testing but the unit tests now fulfill this role.
Since backup is the only code concerned with the primary flag, move the code into the backup module.
We don't have any cross-version testing but this change was tested manually with the most recent version of pgBackRest to make sure it was tolerant of the missing primary info. When an older version of pgBackRest loads a newer manifest the primary flag will always be set to false, which is fine since it is not used.
BackupJobData has several members that backupProcessQueue() needs so it is more efficient to use them rather than passing them separately or getting them from the configuration.
Coverity pointed out that -1 could be passed to lseek() (added in a79034ae) after a file failed to open because it is missing. Overall it seems simpler to enclose the success code in an else block to prevent any repeats of this mistake in the future.
This was not an active bug because there are currently no cases where we do read offsets in a file that is allowed to be missing.
Also remove the result flag since it is easier to just check that the file descriptor is valid.
Updating the manifest this way was not a great idea because it broke abstraction for the object. This meant certain changes to the interface and internals were not possible because the code was modifying internal manifest data.
Instead track the user replacements entirely in the restore module.
This also has the benefit of eliminating a pass over the manifest path/file/link lists.
AWS S3 integrates with AWS Key Management Service (AWS KMS) to provide server side encryption of S3 objects. This integration protects objects under encryption keys that never leave AWS KMS unencrypted.
The range feature allows reading out an arbitrary chunk of a file and will be important for efficient small file support.
Now that all drivers are required to support ranges remove the storageFeatureLimitRead feature flag that was implemented only by the Posix driver.
Do the replacement anywhere cfgOptionGroupIdxToKey() is being used to construct a group name in a message. cfgOptionGroupName() is better for this case since it also includes the name of the group so that it does not need to be repeated in each message.
Functionality to copy from IoRead to IoWrite is frequently used so centralize it. This also simplifies coverage testing in places where a loop was required before.
The backup LSNs are useful for performing LSN-based PITR. LSNs will not be displayed in the general text output (without --set) because they are probably not useful enough to deserve their own line.
There is no evidence that users need 8.3/8.4 anymore but it does cost us in terms of development and testing, especially now that we have a number of new backup/restore features planned.
It seems to make sense to remove this support now. If there are users who need to use/migrate from these versions they can use an older version of pgBackRest.
