Checking that pd_upper == 0 is not enough since this field may be corrupted. Still use pd_upper as a quick check, but when it is zero proceed to check the rest of the page to ensure it is also all zeroes.
Rather than attempting to filter page checksum failures by LSN, just retry when there is a page checksum failure. If the page has not changed since the last read report it as an error. If the page has changed, then PostgreSQL must be modifying the page so we can ignore the error because a full page write (and possibly updates) will be in the WAL.
Also remove tests made redundant by the test merge in b4897077.
Retry a WAL segment that was previously reported as missing by the archive-get command. This prevents notifications in the spool path from a prior restore from being used and possibly causing a recovery failure if consistency has not been reached.
Disabling this option allows PostgreSQL to more reliably recognize when the end of the WAL in the archive has been reached, which permits it to switch over to streaming from the primary. With retries enabled, a steady stream of WAL being archived will cause PostgreSQL to continue getting WAL from the archive rather than switch to streaming.
When disabling this option it is important to ensure that the spool path for the stanza is empty. The restore command does this automatically if the spool path is configured at restore time. Otherwise, it is up to the user to ensure the spool path is empty.
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.
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.
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.
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.
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.
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.
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.
This leak was caused by the file descriptor variable getting clobbered after a long jump. Mark it as volatile to fix.
Testing this is a bit complex because the issue only happens in optimized builds, if at all. Put the test into the performance suite, which is always optimized, until a better idea presents itself.
If a path/file was remapped to a link using either --link-map or --link-all there would be no affect if the path/file already existed. If a link existed it would be properly updated and converting a link to a path/file also worked.
The issue happened during delta cleanup, which failed to check if the existing path/file had been remapped to a link.
Add checks for newly mapped path/file links and remove the old path/file we required.
This was previously a warning but the warning is easy to miss so a lot of time may be lost restoring and recovering a backup that will not hit the target.
Since this is technically a breaking change, add an "important note" about the change to the release.
Fail the backup if a cluster stops or the standby is promoted. Previously, shutting down the primary would cause an error but it was not detected until the end of the backup. Now the error will happen sooner and a promotion on the standby will also cause an error.
SIGHUP allows the configuration to be reloaded. Note that the configuration will not be updated in child processes that have already started.
SIGTERM terminates the server process gracefully and sends SIGTERM to all child processes. This also gives the tests an easy way to stop the server.
Add the following checks:
* Checkpoint is updated in pg_control after pg_start_backup(). This helps ensure that PostgreSQL and pgBackRest have a consistent view of the storage and that PGDATA paths match.
* Timeline of backup start WAL file matches pg_control. Hard to see how this one could get hit, but we have the power...
* Standby is on the same timeline as the primary. If not, this standby is not following the primary.
* Last standby checkpoint is not greater than the backup checkpoint. If so, this standby is not following the primary.
This also requires some additional plumbing to read/write timeline/checkpoint from pg_control and parse timelines from WAL filenames. There were some changes in the backup tests caused by the fact that pg_control now has different contents for each backup.
The check to ensure that the required checkpoint was reached on the standby should also be updated to use pg_control (it currently uses pg_control_checkpoint()), but that requires non-trivial changes to the test harness and will need to wait.
Previously the archive was only checked at the end of the backup to ensure all WAL required to make the backup consistent was present. The problem was that if archiving was not functioning then the backup had to complete before the user found out, which could be a while if the database was large enough.
Add an archive check immediately after backup start so failures are reported earlier.
The trick is to determine which WAL to check. If the repo is new there may not be any WAL in it and pg_start_backup() will not switch the WAL segment if it is empty. These are both likely scenarios when setting up and/or testing pgBackRest.
If the WAL segment is switched by pg_start_backup(), then check the archive for the segment that was detected prior to backup start. This should be common on normal running clusters with regular activity. Note that this might not be the segment immediately prior to the backup start segment if WAL volume is high.
If pg_start_backup() did not switch the WAL then we can force a switch on PostgreSQL >= 9.3 by creating a restore point. In that case the WAL to check will be the backup start WAL. This is most likely to happen on idle systems, during testing, or immediately after a repo switch.
An advantage of this approach other than earlier notification is that the backup directory will not be created so no resume will be attempted on the next backup.
Note that some additional churn was created in backup.c because the load of archive.info needs to be done earlier.
Size option default and allowed values were displayed in bytes, which was confusing for the user.
This also lays the groundwork for adding units to time options.
Move option parsing functions into a common module so they can be used from the build module.
The error handler used a loop to process try, catch, and finally blocks. This worked fine but static analysis tools like Coverity did not understand that the finally block would always run and so there were false positives about double-free, unfreed resource, etc.
This implementation removes the loop, which simplifies everything, and makes it clear that the finally block will always run. This cuts down on Coverity false positives.
This implementation also catches lack of coverage on empty catch blocks so a few test fixes were committed separately in d74fe7a.
A small refactor in backup.c is required because gcc 10.3.1 on Fedora 33 complains that the reason variable may be used uninitialized. It's not clear why this is the case, but reducing the scope of the TRY block fixes the issue.
Rather the converting String to StringIds at runtime, store defaults in StringId format in parse.auto.c and convert user input to StringId during parsing.
The compress-type, repo-type and log-level-* options have allow lists, which means it is more efficient to treat them as StringIds.
For compress-type and log-level-* also update the functions that convert them to enums.
The strIdFrom*() forced the caller to pick an encoding, which led to a number of TRY...CATCH blocks in the code. In practice the caller does not care which encoding is used as long as the string is valid for some encoding.
Update the strIdFrom*() function to try all possible encodings and only throw an error when the string is not valid for any of them.
The variants were needed to easily serialize configurations for the Perl code.
Unions are more efficient and will allow us to add new types that are not supported by variants, e.g. StringId.
A stanza name like global_stanza was not allowed because the code was not selective enough about how a global section should be formatted.
Update the config parser to correctly recognize global sections.
Currently link-map only allows links that exist in the backup manifest to be remapped to a new destination.
Allow link-map to create a new link as long as a valid path/file from the backup is referenced.
The local process will retry jobs (e.g. backup file) but after a certain number of failures gives up. Previously, the last error was reported but generally the first error is far more valuable. The last error is likely to be a cascade failure such as the protocol being out of sync.
Report the first error (and stack trace) and append the retry errors to the first error without stack trace information.
"error list" makes it clearer that other errors may be reported. For example, if checksum-page is true in the manifest but no checksum-page-error list is provided then the error is in alignment, i.e. the file size is not a multiple of the page size, with allowances made for a valid-looking partial page at the end of the file.
It is still not possible to differentiate between alignment and page checksum errors in the output but this will be addressed in a future commit.
The prior limitations were based on using getopt_long() to parse command-line options, which required a static list of allowed options. Setting index max too high bloated the binary unacceptably. 45a4e80 replaced the functionality of getopt_long() but the static list remained.
Improve cfgParseOption() to use available option data and remove the need for a static list. This also allows the option deprecations to be represented more compactly.
Index max is still capped at 256 because a large enough index could cause parseOptionIdxValue() to run out of memory since it allocates a static list based on the highest index found. If that function were improved with a map of found index values then index max could be set to UINT64_MAX.
Note that deprecations no longer set an index max or define whether reset is valid. These were space-saving measures which are no longer required. This means that indexed deprecated options will also be valid up to 256 and always allow reset, but it doesn't seem worth additional code to limit this behavior.
cfgParseOptionId() is no longer needed because calling cfgParseOption() with .ignoreMissingIndex = true duplicates the functionality of cfgParseOptionId(). This leads to some simplification in the help code.
The backup size was a bit off because it did not include any files (e.g. backup_label, WAL files) that were added to the manifest after the main copy. To fix this move the log message to the very end of the backup.
Add size/file total log message to restore since it did not exist before.
Remove the "Automatic Stop Option" section since it only applies to PostgreSQL <= 9.6, which will soon be EOL. Since we no longer build the user guide for PostgreSQL < 10 this section was no longer being tested. The stop-auto option is still documented in the reference.
Move the "Fast Start Option" to "Quick Start - Perform Backup". This is a commonly-used option so it makes sense to mention it earlier. This also makes the backups run more quickly. In the worst case, backups in "Quick Start - Perform Backup" could take minutes to start
Move the "Archive Timeout" section to "Quick Start - Perform Backup" since it is the last section in "Backup".
Options for other repos can cause conflicts and should never be used. Each remote can address exactly one repo or pg cluster.
Also fix an outdated comment.
pg1 was incorrectly used instead of {[host-pg1]} which meant the wrong host name was displayed.
Also, the install block was installing packages to the build host no matter which host was specified.
If the test path is inside the repo path then it can cause strange issues during testing because the entire repo path is duplicated into the test path so that all tests see a consistent view of the repo.
Another solution might be to pick a better test path name and exclude it from the rsync, but this fix at least addresses the immediate issue.
Multi-part upload may fail despite returning an HTTP success code. Check for the ETag field in the result and if not present consider the upload to have failed. This will trigger a retry at the local job level.
Links were followed before they were checked for validity so a circular link would send the manifest build into endless recursion leading to a crash. Fix by moving the recursion after the link check.
Note that this issue has existed since the C migration and was not introduced by the refactor in eba013b.
Switch from JSON-based to binary protocol for communicating with local and remote process. The pack type is used to implement the binary protocol.
There are a number advantages:
* The pack type is more compact than JSON and are more efficient to render/parse.
* Packs are more strictly typed than JSON.
* Each protocol message is written entirely within ProtocolServer/ProtocolClient so is less likely to get interrupted by an error and leave the protocol in a bad state.
* There is no limit on message size. Previously this was limited by buffer size without a custom implementation, as was done for read/writing files.
Some cruft from the Perl days was removed, specifically allowing NULL messages and stack traces. This is no longer possible in C.
There is room for improvement here, in particular locking down the allowed sequence of protocol messages and building a state machine to enforce it. This will be useful for resetting the protocol when it gets in a bad state.
It is better to clear errors after the catch block completes rather than leave them set until the next error. This also make is possible to tell when a error is currently being handled, which a function further down the stack might use to modify its behavior. Currently this is only useful in testing, but clearing the error seems like a good idea in general.
Two places used errors outside the CATCH() block. Mem context cleanup now uses a FINALLY() which is a better implementation anyway. The error handling in main() now calls exitSafe() from withing the CATCH() block.
Since the pack type was stored in 4 bits, only 15 values were allowed (0 was reserved).
Allow virtually unlimited types by storing type info in a base-128 encoded integer following the tag when the type bits in the tag are set to 0xF.
Also separate the type IDs used in the pack (PackTypeMap) from those presented to the user (PackType). The prior PackType enum exposed implementation details to the user, e.g. pckTypeUnknown.
There are no code changes from PostgreSQL 13 so simply add the new version.
Add CATALOG_VERSION_NO_MAX to allow the catalog version to "float" during the PostgreSQL beta/rc period so new pgBackRest versions are not required when the catalog version changes.
Update the integration tests to handle new PostgreSQL startup messages.
manifestLinkCheck() was pretty inefficient so large numbers of links caused it to use a lot of memory and eventually crash. This is a more efficient implementation which runs O(nlogn) and uses far less memory.
Checking for duplicate file links has been added, which represents a change in behavior, but hopefully a good one.
The user guide was split primarily to provide documentation for the stop-auto option in PostgreSQL <= 9.5. Now that 9.5 is EOL there does not seem to be a good reason to generate an extra user guide. The stop-auto function is still documented in the reference.
Leave the stop-auto documentation in the user guide in case we want to manually generate documentation for older versions.
Also rename centos to rhel for most identifiers since that is the core platform we are building for, similar to how we label 'debian' builds even though we generally use Ubuntu. With CentOS set to become an upstream for RHEL later this year, we'll likely need to pick a new test distribution, perhaps Rocky Linux if that gets off the ground.
getopt_long() requires an exhaustive list of all possible options that may be found on the command line. Because of the way options are indexed (e.g. repo1-4, pg1-8) optionList[] has 827 entries and we have kept it small by curtailing the maximum indexes very severely. Another issue is that getopt_long() scans the array sequentially so parsing gets slower as the index maximums increase.
Replace getopt_long() with a custom implementation that behaves the same but allows options to be parsed with a function instead of using optionList[]. This commit leaves the list in place in order to focus on the getopt_long() replacement, but cfgParseOption() could be replaced with a more efficient implementation that removes the need for optionList[].
This implementation also fixes an issue where invalid options were misreported in the error message if they only had one dash, e.g. -config. This seems to have been some kind of problem in getopt_long(), but no investigation was done since the new implementation fixes it.
Tests were added at 0825428, 2b8d2da, 34dd663, and 384f247 to check that previously untested getopt_long() behavior doesn't change.
Remove stanza archive spool path so existing files do not interfere with the new cluster. For instance, old archive-push acknowledgements could cause a new cluster to skip archiving. This should not happen if a new timeline is selected but better to be safe. Missing stanza spool paths are ignored.
Also add new path expression STORAGE_SPOOL_ARCHIVE to easily access this path.
When running on a GCE instance the authentication token can be pulled directly from the instance metadata. This is configured with repo-gcs-key-type=auto.
In a separate commit (26fefa6), move the code that parses the token response into a separate function, storageGcsAuthToken(), since it is now needed by two key types. This drastically improves the readability of the main commit.
If an ok file (which indicates the WAL segment was not found) is present on the first iteration of the loop then remove it and spawn the async process to retry. This action also resets the queue.
Also error if no response is received from the async process rather than returning not found. PostgreSQL will respond the same either way, but this allows us to determine when something is going wrong with the async process.
Update archiveAsyncStatus() to allow warnings to be suppressed. It is better to retry if no WAL segment was found before warning because the warning might be stale.
Convert most of the remaining options that benefit from being StringIds. Since all the command modules can include config.h directly it makes sense to auto-generate these values instead of manually creating an enum for each one.
For the time being StringIds are not being auto-generated because the StringId code does not exist in Perl. However, the *_Z zero-terminated constants for each allowed option value are now auto-generated.
Allows removal of backupType()/backupTypeStr() and improves debug logging of the enum.
Move BackupType enum and string constants to info/infoBackup.h so they are available to more modules. Also convert InfoBackup to use BackupType instead of a String.
Centralize the formatting of the configuration value for display to the user or passing on a command line.
For the new functions, if the value was set by the user via the command line, config, etc., then that exact value will be displayed. This makes it easier for the user to recognize the value and saves having to format it into something reasonable, especially for time and size option types.
Note that cfgOptTypeHash and cfgOptTypeList option types are not supported by these functions, but they are generally not displayed to the user as a whole.
This also fixes a bug in config/load.c where time values where not being formatted correctly in an error message.
Use StringIds for the storage types (e.g. STORAGE_S3_TYPE) and configuration settings, e.g. cfgOptS3KeyType.
Also add new config functions and harness config functions to support StringIds.
There is no need to write the file atomically (e.g. via a temp file on Posix) because checksums are tested on resume after a failed backup. The path does not need be synced for each file because all paths are synced at the end of the backup.
This functionality was not lost during the migration -- it never existed in the Perl code, though these settings are used in restore. See 59f1353 where backupFile() was migrated to C.
Fix the segfault when getting help for an internal option is requested by adding help for all internal options that are valid for a default command role.
Also print warnings about internal options in code rather than putting in each command/option description.
The remotes are no longer needed in the main process after the manifest is loaded. If the restore is long enough the connection will timeout and WARN at the end of the restore. This is harmless for the restore but distracting for the user.
To prevent this, free the remotes once they are no longer needed.
Getting help for a valid option that was invalid for the command would segfault.
Add a check to ensure the option is valid for the command's default role.
Enabled by default, this option checks the WAL header against the PostgreSQL version and system identifier to ensure that the WAL is being copied to the correct stanza. This is in addition to checking pg_control against the stanza and verifying that WAL is being copied from the same PostgreSQL data directory where pg_control is located.
Therefore, disabling this check is fairly safe but should only be done when required, e.g. if the WAL is encrypted.
3b8f0ef missed some cases that could cause archive-push to fail:
* Checking archive info.
* Checking to see if a WAL segment already exists.
These cases are now handled so archive-push can succeed on any valid repos.
The content-length header was being signed since it was the only header that didn't need to be and it seemed simpler just to sign it as well. Also, the S3 documentation encourages signing as many headers as possible to avoid tampering.
However, some proxies munge this header causing authentication failure, so skip signing content-length.
Make protocol handlers have one function per command. This allows the logic of finding the handler to be in ProtocolServer, isolates each command to a function, and removes the need to test the "not found" condition for each handler.
S3 returns 200 for HEAD / which indicates it is a file but does not return the expected headers which causes an error.
Rather than fix this for S3, just automatically return / as not existing for any storage that does not support paths.
Also add some defensive checks to prevent this from generating a segfault if it happens again.
Recovery may error unless --type=immediate is specified. This is because after consistency is reached PostgreSQL will flag zeroed pages as errors even for a full-page write.
For PostgreSQL ≥ 13 the ignore_invalid_pages setting may be used to ignore invalid pages. In this case it is important to check the logs after recovery to ensure that no invalid pages were reported in the selected databases.
It is best if the archive-push and backup commands have the same compress-type (e.g. lz4) when using archive-copy. Otherwise, the WAL segments will need to be recompressed with the compress-type used by the backup, which can be fairly expensive depending on how much WAL was generated during the backup.
Up to four repositories may be configured. A potential benefit is the ability to have a local repository for fast restores and a remote repository for redundancy.
Some commands, e.g. stanza-create/stanza-update, will automatically work with all configured repositories while others, e.g. stanza-delete, will require a repository to be specified using the repo option. See the command reference for details on which commands require the repository to be specified.
Note that the repo option is not required when only repo1 is configured in order to maintain backward compatibility. However, the repo option is required when a single repo is configured as, e.g. repo2. This is to prevent command breakage if a new repository is added later.
The archive-push command will always push WAL to the archive in all configured repositories but backups will need to be scheduled individually for each repository. In many cases this is desirable since backup types and retention will vary by repository. Likewise, restores must specify a repository. It is generally better to specify a repository for restores that has low latency/cost even if that means more recovery time. Only restore testing can determine which repository will be most efficient.
For single repository configurations there should be no change in behavior.
The following options are renamed as specified:
repo1-azure-ca-file -> repo1-storage-ca-file
repo1-azure-ca-path -> repo1-storage-ca-path
repo1-azure-host -> repo1-storage-host
repo1-azure-port -> repo1-storage-port
repo1-azure-verify-tls -> repo1-storage-verify-tls
repo1-s3-ca-file -> repo1-storage-ca-file
repo1-s3-ca-path -> repo1-storage-ca-path
repo1-s3-host -> repo1-storage-host
repo1-s3-port -> repo1-storage-port
repo1-s3-verify-tls -> repo1-storage-verify-tls
The old option names (e.g. repo1-s3-port) will continue to work for repo1, but repo2, etc. will require the new names.
Option warnings will cause the async process to fail because a warning is logged but stdout is closed so the process aborts.
This bug has existed for quite some time, but it was made worse by abb8ebe because now the async role can have different valid options than the default role. Previously at least a warning would be emitted before the async process died.
Fix this by only allowing warnings for the default role. Warnings were already suppressed for local and remote roles so the logic already exists.
The destination buffer on the stack was not large enough to contain the zero-terminating character.
Increase the buffer size and add an assertion to prevent regressions.
Found on arm64 running musl libc. Other architectures and glibc do not seem to be affected though it is clearly a bug.
Check that archive files exist in the main process instead of the local process. This means that the archive.info file only needs to be loaded once per execution rather than once per file to get.
Stop looking when a file is missing or in error. PostgreSQL will never request anything past the missing file so there is no point in getting them. This also reduces "unable to find" logging in the async process.
Cache results of storageList() when looking for multiple files to reduce storage I/O.
Look for all requested archive files in the archive-id where the first file is found. They may not all be there, but this reduces the number of list calls. If subsequent files are in another archive id they will be found on the next archive-get call.
If files other than backup.manifest.copy were left in a backup path by a prior resume then the next resume would skip the backup rather than removing it. Since the backup path still existed, it would be found during backup label generation and cause an error if it appeared to be later than the new backup label. This occurred if the skipped backup was full.
The error was only likely on object stores such as S3 because of the order of file deletion. Posix file systems delete from the bottom up because directories containing files cannot be deleted. Object stores do not have directories so files are deleted in whatever order they are provided by the list command. However, the issue can be reproduced on a Posix file system by manually deleting backup.manifest.copy from a resumable backup path.
Fix the issue by removing the resumable backup if it has no manifest files. Also add a new warning message for this condition.
Note that this issue could be resolved by running expire or a new full backup.
These options specify the number of local worker job retries and the retry interval after one immediate retry.
There is some value in allowing retries to be specified by the user but for the most part these options are for suppressing retries during testing, which can save a lot of time. The bug introduced in d1d25c7 and fixed in 8b86d5e also suggests it is better not to use retries in tests.
Remove the default delayed retries for archive-get/archive-push, leaving only the immediate retry. These commands are retried by PostgreSQL so it doesn't make sense to do too many retries internally.
These options are currently internal.
Building on 23f5712, limit option validity by role. This is mostly for options that weren't needed for certain roles but were harmless. However, the upcoming multi repository functionality requires the granularity implemented here.
The remote role benefits since host options can automatically excluded when building the options. Also, many options that are only required for the default role (e.g. repo-retention-full) no longer need to be passed in tests for other roles.
Validity by command was not granular enough so numerous options needed be marked internal so users would not stumble across them. Options were also needlessly being passed to roles that had no use for them.
Introduce per-role validity lists that depend on what roles are valid per command. Also add a check to ensure that only valid roles are used with a command.
This commit adds the functionality but does not introduce any new behavior, i.e. all options are valid for all roles that the command is valid for. A subsequent commit will introduce the new role restrictions to make the changes easier to audit.
The pack type is an architecture-independent format for serializing data compactly, inspired by ProtocolBuffers and Avro.
Also add ioReadSmall(), which is optimized for small binary reads, similar to ioReadLineParam().
The C code does not use doubles to represent seconds like the Perl code did so time can be represented as an integer which reduces the number of data types that config has to understand.
Also remove Variant doubles since they are no longer used.
Note that not all double code was removed since we still need to display times to the user in seconds and it is possible for the times to be fractional. In the future this will likely be simplified by storing the original user input and using that value when the time needs to be displayed.
Improve locking on remote processes by introducing an exec-id that is unique to the main process and passed to all remote processes. This allows the remote processes to determine if a lock is held by a remote from the same main process. If so, the lock is allowed.
The exec-id is also useful for associating remote logs with main logs for debugging purposes.
lstRemoveIdx(list, 0) resulted in the entire list being moved down to the first position which could take a long time for big lists. This is a common pattern in backup/restore when processing file queues.
Instead simply move the list pointer up when first item is removed. Then on insert check if there is space at the beginning when there is no longer space at the end and do the move then. This way if a list is built and then drained without any new inserts then no move is required.
There were a number of places in the code where "hostId" was used, but hostId is just the option group index + 1 so this led to a lot of +1 and -1 to convert the id to an index and vice versa.
Instead just use the zero based index wherever possible. This is pretty much everywhere except when the host-id option is read or set, or where a message is being formatted for the user.
Also fix a bug in protocolRemoteParam() where remotes spawned from the main process could get process ids that were not 0. Only the locals should spawn remotes with process id > 0. This seems to have been harmless since the process id is only a label, but it could be confusing when debugging.
iniLoad() was trimming lines which meant that a leading space would not pass checksum validation when a manifest was reloaded. Remove the trims since files we write should never contain extraneous spaces. This further diverges the format for the functions that read conf files (e.g. pgbackrest.conf) and those that read info (e.g. manifest) files.
While we are at it also allow [ and # as initial characters. # was reserved for comments but we never put comments into info files. [ denotes a section but we can get around this by never allowing arrays as values in info files, so if a line ends in ] it must be a section. This is currently the case but enforce it by adding an assert to info/info.c.
Instead of using memmove() to manage the internal output buffer for every small read, track the current buffer position and only move data when the small read cannot be satisfied and more data is needed.
Group related options together so operations (e.g. valid, test, index total) can be performed on all options in the group.
Previously, options at the top of the hierarchy of the related options were used to do these tests. This was prone to error as option relationships changed and it was not always clear which option (or options) should be used.
This means the same text will appear in both places, which should make it easier to find.
Also update the link code to allow both page and section to be specified rather than only one or the other.
Update the documentation to explicitly state that versions must match across hosts when running remotely.
Add a hint to the protocol version mismatch error to help the user identify the problem.
Previously, catalog versions were fixed for all versions which made maintaining the catalog versions during PostgreSQL beta and release candidate cycles very painful. A version of pgBackRest which was functionally compatible was rendered useless by a catalog version bump in PostgreSQL.
Instead use only the control version to identify a PostgreSQL version when possible. Some older versions require a catalog version to positively identify a PostgreSQL version, so include them when required.
Since the catalog number is required to work with tablespaces it will need to be stored. There's already a copy of it in backup.info so use that (even though we have been ignoring it in the C versions).
This condition used to give a not-very-clear error which we have been intending to improve. But in the meantime the changes in fbff299 resulted in a segfault for this condition instead because the data_directory was assumed to be non-NULL.
Fix this by explicitly throwing an error with hints when any row in pg_settings cannot be selected.
This file is created by pg_basebackup so might be in the data directory if the cluster was restored from a pg_basebackup backup. Also exclude backup_manifest.tmp since it is possible to find that in the backup directory.
Improve the wording of the error message and add a hint to make it clearer what is wrong and how the user can fix it.
Also change the assert to a regular error since this is not an internal error.
If the callback never returned any jobs then protocolParallelDone() would never be true. The reason is that the done state was being set in protocolParallelResult(), which never gets called if there are no results.
Calling protocolParallelResult() doesn't make much sense in this case so instead move the done logic to protocolParallelDone().
For current usage of ProtocolParallel we ensure there are jobs before processing so this is not a live issue, but the new behavior is required for future development.
When restoring a cluster that will be promoted but is not intended to be the new primary, it is important to disable archiving to avoid polluting the repository with useless WAL. This option makes disabling archiving a bit easier.
Automatically retrieve the role and temporary credentials for S3 when the AWS instance is associated with an IAM role. Credentials are automatically updated when they are <= 5 minutes from expiring.
Basic configuration is to set repo1-s3-key-type=auto. repo1-s3-role can be used to set a specific role, otherwise it will be retrieved automatically.
David Steele