VERSION
This document describes version 1.1.80 of Rinci::Transaction (from Perl distribution Rinci), released on 2016-05-10.SPECIFICATION VERSION
Rinci 1.1, protocol version 2
SPECIFICATION
This document describes a transactional system based on functions, where several function calls participate in a single transaction. This transactional system has the following properties:-
Client/server architecture
Transaction can be performed over Riap. Client can start more than one active transaction on the server. Each transaction-management request and the function calls are requested separately (each one is a separate Riap request).
For more details on this, see Riap::Transaction.
-
Undo/redo
Committed transactions are still recorded in the database along with its undo information. Client can request to undo/redo the transactions. Thus the system is also an undo/redo system.
-
Relies on the functions for reliability/ACID properties
Server or framework provides the transaction manager (TM), but each function acts as the resource manager (RM). It is the responsibility of the functions to maintain ACID properties while modifying resources. For best results, each function should be written carefully and tested extensively, and utilize a real, robust RM (like an RDBMS to store data or a transactional filesystem layer to read/modify files). In the absence of a real RM, some ACID properties like isolation and consistency might be compromised. For example: one transaction TX1 modifies a file in an ordinary (i.e. non-transactional) filesystem. Another transaction TX2 can see TX1's modification in the middle of uncommitted transaction (violates isolation principle).
How transaction works
The basic idea is that actions are performed by function calls. For each action, TM will call the function twice. First for getting undo information, and second for actually performing the action. The undo information can be used to perform rollback, undo, and redo. All functions performing actions in the transaction must be able to supply undo information.Function requirements
Functions that participate in transaction must declare their "tx" feature in the metadata. In addition, function must also be idempotent.
features => { ... tx => {v=>2}, idempotent => 1, }
Function must then follow the transaction protocol, described below.
Transaction manager
The transaction manager manages transaction data and performs actions as well as transaction management.For the sake of examples, our TM stores data in a SQL database (like SQLite) with the following tables:
-
tx
id (PK) summary ctime (creation time) commit_time status last_action_id -- in-progress action ID (for tx with status=i), or last -- processed action (for tx with other transient statuses)
-
do_action
id (PK) tx_id (refers to tx(id)) ctime sp (savepoint name, UNIQUE(sp,tx_id)) f (function name) args (arguments, serialized)
-
undo_action
id (PK) tx_id (refers to tx(id)) ctime f (function name) args (arguments, serialized)
Transaction status
A transaction can have one of these statuses. They will be fully explained in the following sections. Statuses having lowercase labels are transient statuses. Statuses having uppercase labels are final statuses.
i (in-progress) a (aborted, pending rollback to R) R (rolled back) C (committed) u (committed, undoing) v (aborted undoing, pending rollback back to C) U (committed, undone) d (committed & undone, redoing) e (aborted redoing, pending rollback back to U) X (unresolvable/error)
Transaction manager initialization
User instantiates TM. TM sets up its data directory and performs cleanup and crash recovery.In cleanup, TM purges unneeded data, like data for rolled back transactions or committed transactions that have been around for too long.
In crash recovery, TM looks at all crashed transactions and resolves them (either by performing rollback or roll forward). Crashed transactions are in-progress ("i") transactions that have an in-progress action, or transactions having one of these statuses (all the other transient statuses): "a", "u", "v", "d", "e". Crash recovery will be explained in the following sections below.
TM also can perform rollback for in-progress transactions that have been around for too long without being committed or rolled back.
Starting transaction
User invokes "$tm->begin(tx_id => $tx_id)", providing a unique transaction ID $tx_id as identifier for the transaction. $tx_id is an arbitrary string with a length between 1 and 200 characters. User can also supply "summary", a textual description for the transaction. It should not be longer than 1024 characters. TM will create an entry for the transaction in its journal:
BEGIN; INSERT INTO tx (id,summary,ctime,status) VALUES ($tx_id,$summary,$now,'i') COMMIT;
As can be seen, initial transaction status is "i" (in-progress).
Upon success, TM must return status 200. If transaction with that ID already exists, TM must return status 409, unless when the existing transaction is still on-going, in which case TM should just return 200. TM must return 400 if no $tx_id is given. TM can also return status 412 if there are already too many transactions being started, either globally on the server or for the particular client.
Performing action
1) User performs action by invoking "$tm->action(f=>$fname, args=>$args)" one or several times. Transaction status must be "i". TM will first check whether function exists and supports transaction. If function does not exist, or does not support transaction, TM must return status 412.2) TM records this action in its journal:
BEGIN; INSERT INTO action (tx_id,ctime,f,args) VALUES ($tx_id,$now,$fname,JSON($args)); -- $act_id UPDATE tx SET last_action_id=$act_id WHERE id=$tx_id; COMMIT;
3) TM requests state checking and undo information to function, by calling the function using the arguments $args and a special argument "-tx_action=>'check_state'". In addition TM also passes "-tx_v => N" (the protocol version) and "-tx_action_id => UUID" (a unique identifier to link between this call and the 'fix_state' call later).
There are 3 possible states that the function must decide which we are in:
-
fixed
This is the final, desired state. When we are already in a fixed state, function must return status 304 (nothing to do). TM will then skip calling the function the second time to fix state, since there is nothing to fix. For example:
[304, "File $path already exists"] # e.g., in a create_file() function [304, "User $u already does not exist"] # e.g., in a delete_user() function
-
fixable
This is where the final, desired state has not been reached, but it is possible to reach it. When we are in this state, function must return status 200 with the result metadata "undo_actions". The message should also describe what needs to be fixed.
For example:
[200, "Directory $path needs to be created", undef, {undo_actions => [rmdir => {path=>$path}]}] # e.g. in a mkdir() function [200, "User $u should be created with UID $uid", undef, {undo_actions => [delete_user=>{user=>$u}]}] # e.g. in create_user()
-
unfixable
This is where the final, desired state has not been reached, and it is impossible or inappropriate for the function to fix into the fixed state. This state is used to avoid undoing what was not fixed by the function. If we are in this state, function should return status 412 (precondition failed).
For example:
[412, "Path $path exists but not a symlink"] # e.g. in setup_symlink() [412, "User $u exists but with different UID $cur_uid (needs $uid)"]
If state is unfixable, or function returns other statuses (assumed as failure), TM stops the process and starts a rollback. "$tm->action()" will return with the function's result.
For example, let us use function "My::setup_unix_user()" which can create a Unix user with an empty home directory if the user has not been created. This function utilizes several simpler functions: "My::adduser()" to add entry to /etc/passwd and /etc/shadow, "My::addgroup" to add entry to /etc/group and /etc/gshadow, "My::mkdir" to create directory. Then there are also these functions for the undo actions: "My::deluser" to delete user entry in Unix passwd database, "My::delgroup" to delete group entry in Unix group database, and "My::rmdir" to remove directory.
For "My::adduser", the fixable state is that the user does not exist, the fixed state is that the user exists. For "My::deluser", the fixable state is that user exists (additionally with the same UID as the one created previously), the fixed state is user does not exist, the unfixable state is user exists but with different UID. For "My::addgroup", the fixable state is that group does not exist, the fixed state is that the group exists. For "My::delgroup", the fixable state is that group exists (additionally with the same GID as the one created previously), the fixed state is group does not exist, the unfixable state is group exists but with different GID. For "My::mkdir", the fixable state is path does not exist, the fixed state is directory exists, and unfixable state is path exists but is not a directory. For "My::rmdir", the fixable state is directory exists and empty, the fixed state is path does not exist, the unfixable state is path exists but not a directory or directory is not empty.
The "undo_actions" must be an array containing action information, in reverse order. Each action is a two-element array "[$fname, $args]" where $fname is name of a function (not necessarily the same function) and $args its call arguments.
For example, if user invokes "$tm->action(f=>'My::setup_unix_user', args=>{user=>'bob'})" and user "bob" does not exist yet, function will return:
[200, "OK", undef, {undo_actions=>[ ['My::deluser', {group=>'bob'}], ['My::delgroup', {group=>'bob'}], ['My::rmdir', {path=>'/home/bob'}], }]
4) TM records these undo actions in its journal:
BEGIN; INSERT INTO undo_action (tx_id,ctime,action_id,f,args) VALUES ($tx_id,$now,$act_id,'My::deluser','{"group":"bob"}'); -- # $uact_id1 INSERT INTO undo_action (tx_id,ctime,action_id,f,args) VALUES ($tx_id,$now,$act_id,'My::delgroup','{"user":"bob"}'); -- # $uact_id2 INSERT INTO undo_action (tx_id,ctime,action_id,f,args) VALUES ($tx_id,$now,$act_id,'My::rmdir','{"path":"/home/bob"}'); -- # $uact_id3 COMMIT;
5) If we are in fixed state, this step is skipped.
If we are in fixable state, TM calls function the second time, this time with "-tx_action => 'fix_state'". TM also passes "-tx_v" and "-tx_action_id" with the same value as the one passed previously during the 'check_state' call. Function must perform action to fix the state into the fixed state. In our example, "setup_unix_user()" should create user and group "bob", and creates an empty directory "/home/bob".
Function must return status 200 on success. Other status will be interpreted as failure, in which case TM will stop the process and starts rollback. "$tm->action()" will return with the function's result.
Note: During the 'check_state' phase in step 3, function can also optionally return "do_actions" in its result metadata, for example:
[200, "OK", undef, {do_actions=>[ ['My::adduser', {group=>'bob'}], ['My::addgroup', {group=>'bob'}], ['My::mkdir', {path=>'/home/bob'}], undo_actions=>[ ['My::deluser', {group=>'bob'}], ['My::delgroup', {group=>'bob'}], ['My::rmdir', {path=>'/home/bob'}], }]
In this case, instead of calling function the second time, TM will just call the actions provided by the function, using a nested "$tm->action(actions => $do_actions)". Step 4 will be skipped since each do action will provide its own undo actions.
6) If 'fix_state' phase in step 5 succeeds, the action is finished. TM marks this:
BEGIN; UPDATE tx SET last_action_id=NULL WHERE id=$tx_id; COMMIT;
TM is ready to process another action.
Crash recovery
Recovery rolls back interrupted in-progress transaction. See ``Rollback of in-progress (status i) transaction'' for more details.
If crash happens after step 1, transaction will not be marked as crash since "last_action_id" has not been set and no recovery is necessary.
If crash happens after step 2 until 5, recovery will be performed by rollback. Details of rollback is explained in ``Rollback of in-progress (status i) transaction''.
If crash happens after step 6, transaction will not be marked as crash since "last_action_id" is already unset and no recovery is necessary.
Commit
To commit transaction, user invokes "$tm->commit()". Transaction status must be "i" or "a". If transaction status is "a", transaction must be rolled back instead.TM will mark the transaction status as "C" (committed) and delete all entries in the "do_action" table since they are no longer needed:
BEGIN; UPDATE tx SET status='C' WHERE id=$tx_id; DELETE FROM do_action WHERE tx_id=$tx_id; COMMIT;
TM still stores the "undo_actions" entries for some time, to allow undo (and redo) of transactions.
If transaction status is "a", transaction should be rolled back instead of committed.
Transaction status progress:
i -> C
Rollback of in-progress (status i) transaction
If an action fails, or some other error happens, rollback will be performed by TM. Rollback can also be started by user using "$tm->rollback". TM marks transaction status to "a" (aborted). This will prevent other clients trying to add new actions to this transaction, since aborted transaction can longer accept new actions, it can only be rolled back.TM will then perform undo for each function, in reverse order, using the undo actions previously recorded in "undo_action" table. The process is similar to performing action, except that:
-
After rollback succeeds, transaction status is changed to "R"
"R" means rolled back. These transactions can be discarded by the next cleanup process.
-
Undo actions are not recorded
Since we do not rollback from the rollback process, but continue it. TM still calls function twice for each action (check_state + fix_state), but do not bother to record the undo actions returned by function in the check_state phase to its database.
-
Failure in rollback step will mark transaction status as "X"
"X" means inconsistent/error. Transactions left in this state are probably half-done and thus inconsistent. We give up on these transactions and the next cleanup process can discard them.
(TODO: Should there be an option to continue to the next action anyway? But this is not necessarily more robust or correct.)
Transaction status progress:
i -> a -> R # successful rollback i -> a -> X # failed rollback
Example. Continuing our previous example, in the "setup_unix_user(user=>'bob')" action, there are 3 actions involved:
['My::adduser', {group=>'bob'}] ['My::addgroup', {group=>'bob'}] ['My::mkdir', {path=>'/home/bob'}]
Suppose action 1 and 2 succeed, and the following undo actions have been recorded in "undo_action":
['My::deluser', {group=>'bob'}] # recorded with ID $ucall_id1 ['My::delgroup', {group=>'bob'}] # recorded with ID $ucall_id2
Suppose action 3 fails with status 500 (e.g. permission denied) and thus rollback is started. The following is the steps that happen during rollback. Actions will be processed in reverse order: $ucall_id2, $ucall_id1.
1) TM marks transaction status to aborted:
BEGIN; UPDATE tx SET status='a', last_action_id=NULL WHERE id=$tx_id; COMMIT;
TM performs action "My::delgroup".
2a) TM calls "My::delgroup()" the first time with "-tx_action => 'check_state'". TM also passes "-tx_is_rollback => 1" for informative purposes (some function can utilize this information to behave more robust, for example, to avoid failing the rollback process). TM does not record the "undo_actions" metadata returned, but observes the "do_actions".
If function returns 304, step 2b is skipped and TM moves on to the next action. If function returns 200, TM continues to step 2b. If function returns other statuses, TM assumes rollback failure and marks transaction as "X" and ends the rollback process for this transaction.
2b) TM invokes "My::delgroup()" the second time to perform the action, passing "-tx_action => 'fix_state'" and "-tx_is_rollback => 1". Function sees that group exists (fixable state), deletes it, return status 200.
2c) TM sets transaction's "last_action_id" to $uact_id1 to mark that this action has been processed:
BEGIN; UPDATE tx SET last_action_id=$ucall_id1 WHERE id=$tx_id; COMMIT;
TM then continues to perform action "My::delgroup".
3a) Just like in step 2, TM invokes "My::deluser()" the first time to check state.
3b) TM invokes "My::deluser()" to perform the action. Function sees that user exists (fixable state), deletes it, return status 200.
3c) TM sets transaction's "last_action_id" to $uact_id2 to mark that this action has been processed:
BEGIN; UPDATE tx SET last_action_id=$uact_id2 WHERE id=$tx_id; COMMIT;
4) TM completes the rollback process by setting transaction status to "R".
BEGIN; UPDATE tx SET status='R' WHERE id=$tx_id; COMMIT;
By now the effect of the transaction has been nullified.
* Crash recovery
Recovery continues the interrupted rollback process.
If crash happens after step 1, recovery will continue the rollback process. Rollback of aborted (status a) transaction is exactly the same as rollback of in-progress (status i) transaction, except that "last_action_id" is not reset.
If crash happens after step 2a-2b, "last_action_id" is still unset, so the process resumes at step 2a. TM does not remember whether previously before crash the function has been executed (and cannot remember, the progress of the execution inside the function). This is the reason why function needs to be idempotent, because it is potentially executed twice by TM for the same action. If function has completed deleting the group before crash, "check_state" will return status 304 (fixed) and TM will skip step 2b. If function has not deleted the group before crash, "check_state" will return status 200 (fixable) and TM will execute step 2b.
If crash happens after step 2c/3a-3b, "last_action_id" is set to $uact_id1. Process will resume at step 3a, since $uact_id1 has been marked as done.
If crash happens after step 3c, process will resume at step 4.
If crash happens after step 4, no recovery is necessary since transaction has been rolled back completely.
Undo
TM allows undoing committed transaction, so the transaction system also serves as an undo/redo system.1) User performs undo by invoking "$tm->undo(tx_id => $tx_id)", where $tx_id is the ID of a committed transaction. If $tx_id is not supplied, the client's newest committed transaction is used. TM will first check that transaction status is indeed "C".
2) TM sets transaction status to "u" (undoing):
BEGIN; UPDATE tx SET status='u' WHERE id=$tx_id; COMMIT;
TM then performs actions specified in the "undo_action" table. The process is similar to performing action, except:
-
After undo succeeds, transaction status is changed to "U"
"U" means committed but undone transaction. These transactions can be redone back to status "C".
- Undo actions are recorded in "do_action" table instead of "undo_action"
- Failure in undo step will cause transaction to roll back to status "C"
Transaction status progress:
C -> u -> U # successful undo C -> u -> v -> C # failed undo, rolled back to C
Continuing our previous example, suppose our "setup_unix_user(user=>'bob')" transaction has succeeded and been committed. The "undo_action" table contains these entries:
['My::deluser', {group=>'bob'}] # recorded with ID $uact_id1 ['My::delgroup', {group=>'bob'}] # recorded with ID $uact_id2 ['My::rmdir', {path=>'/home/bob'}] # recorded with ID $uact_id3
Actions will be processed in reverse order: $uact_id3, $uact_id2, $uact_id1.
3a) TM invokes "My::rmdir" the first time with "-tx_action => 'check_state'". If directory has been filled by files/subdirectories, function will return 412 (``Cannot remove home directory, non-empty'') and the undo process fails with this status. If directory exists and is still empty, function will return 200 (fixable state) and process continues.
3b) TM records the "undo_actions" result metadata returned by function to "do_action" table, for redo information.
BEGIN; INSERT INTO do_action (tx_id,ctime,f,args) VALUES ($tx_id,$now,'My::mkdir', '{"path":"/home/bob"}'); # -- $ract_id1 COMMIT;
3c) TM invokes "My::rmdir" the second time with "-tx_action => 'fix_state'". Function deletes directory and return 200.
3d) TM updates "last_action_id" to mark that this action has been processed:
BEGIN; UPDATE tx SET last_action_id=$uact_id3 WHERE id=$tx_id; COMMIT;
TM then continue to $uact_id2.
4a) TM invokes "My::delgroup" the first time with "-tx_action => 'check_state'".
4b) TM records undo_actions:
BEGIN; INSERT INTO do_action (tx_id,ctime,f,args) VALUES ($tx_id,$now,'My::addgroup', '{"group":"bob"}'); # -- $ract_id2 COMMIT;
4c) TM invokes "My::addgroup" the second time with "-tx_action => 'fix_state'". Function sees that group exists, deletes it, and returns 200.
4d) TM updates "last_action_id":
BEGIN; UPDATE tx SET last_action_id=$uact_id2 WHERE id=$tx_id; COMMIT;
TM then continue to $uact_id1.
5a) TM invokes "My::deluser" the first time with "-tx_action => 'check_state'".
5b) TM records undo_actions:
BEGIN; INSERT INTO undo_action (tx_id,ctime,f,args) VALUES ($tx_id,$now,'My::adduser', '{"user":"bob"}'); # -- $ract_id3 COMMIT;
5c) TM invokes "My::adduser" the second time with "-tx_action => 'fix_state'". Function sees that user exists, deletes it, and returns 200.
5d) TM updates "last_action_id":
BEGIN; UPDATE tx SET last_action_id=$uact_id1 WHERE id=$tx_id; COMMIT;
6) TM completes the undo process by setting transaction status to "U":
BEGIN; UPDATE tx SET status='U', last_action_id=NULL WHERE id=$tx_id; COMMIT;
Crash recovery
Recovery rolls back interrupted undoing process so that transaction status is back to "C" (committed). For more details, refer to ``Rollback of undoing (status u) transaction''.
If crash happens before finishing step 2, no recovery is necessary.
If crash happens after step 2-3c, recovery resumes from step 3a since "last_action_id" is still unset. That is why "My::mkdir" needs to be idempotent and can check state, since it is potentially executed (step 3c) twice, before and after recovery.
If crash happens after step 3d-4c, recovery recovery resumes from step 4a since "last_action_id" is set to $uact_id3.
If crash happens after step 4d-5c, recovery resumes from step 5a since "last_action_id" is set to $uact_id2.
If crash happens after step 5d, recovery resumes from step 6.
Rolling back the undoing (status u) transaction
If undo fails in the middle, rollback will happen. TM marks transaction status from "u" to "v", this differentiates between an undo process in progress (in which case recovery should continue it until status is "U") and a failed undo process (in which case recovery should rolls it back to status "C").TM will then perform actions from the "do_action" table. The process is similar to rollback of in-progress (status i) transaction, except that after rollback succeeds, transaction status is set to "C".
If rollback fails, transaction status is set to "X".
Transaction status progress:
u -> v -> C # rollback succeeds u -> v -> X # rollback fails
Crash recovery
Recovery continues the rollback process.
Redo
An undone transaction (status "U") can be redone back to "C". To do this, user invokes "$tm->undo(tx_id => $tx_id)", where $tx_id is the ID of an undone transaction. If $tx_id is not supplied, the client's newest undone transaction is used. TM will first check that transaction status is indeed "U".TM will then set transaction status to "d" (redoing):
BEGIN; UPDATE tx SET status='d' WHERE id=$tx_id; COMMIT;
This will prevent other clients trying to redo the same transaction. TM will then process actions found in "do_action" table, just like when performing normal action.
Transaction status progress:
U -> d -> C
Crash recovery
Recovery rolls back the redoing process. See ``Rolling back a redoing (status d) transaction''.
Rolling back a redoing (status d) transaction
If redo fails in the middle, rollback will happen. TM marks transaction status from "d" to "e" (failed redo). This will differentiate between a redo process in progress (in which case recovery should continue it until status is "C") and a failed redo process (in which case recovery should rolls it back to status "U").TM will perform actions from the "undo_action" table. The process is similar to rollback of an in-progress (status i) transaction, except that after rollback succeeds, transaction status is set to "U".
If rollback fails, TM will set transaction status to "X".
Transaction status progress:
d -> e -> U # rollback succeeds d -> e -> X # rollback fails
Crash recovery
Recovery continues the rollback process.
Cleanup
Cleanup is done at TM startup and at regular intervals. TM should delete (forget) all C and U transactions that are too old, or keep the number of those transactions under a certain limit, according to its settings. As soon as those transactions are deleted, they can no longer be undone/redone, since the undo actions data has been deleted too.The cleanup process also deletes all X transactions, since they cannot be resolved anyway (TODO: perhaps some retry mechanism can be applied, if desired?)
Cleanup process also deletes all R transactions.
Cleanup process can also roll back any transactions with status "i" that have been going for too long without being committed/rolled back.
Savepoint
Basically savepoint is just a label in the "do_action" table.To mark a savepoint, user invokes "$tm->savepoint(sp_id=>$sp_id)" where $sp_id is an arbitrary string from 1-64 characters. It must be unique within the transaction. If the same savepoint is used, the old savepoint is replaced by the new one.
To release (forget) a savepoint, user invokes "$tm->release_savepoint(sp_id=>$sp_id)". It just clears the label in the "do_action" table.
Rollback to a savepoint is just a normal rollback process, except we stop after finishing the undo actions of the corresponding action with the savepoint, and transaction status is set back to "i". If savepoint is unknown (or marked before any action, which is effectively the same), we rollback everything in the transaction.
Discard
User can optionally do a cleanup of her transactions by issuing "$tm->discard(tx_id=>$tx_id)" or "$tm->discard_all". Transactions that can be discarded are those with the final statuses: "C", "U", "X".FAQ
Why is this useful?
The protocol is a pretty generic and simple way to build transactional system, even on heterogenous, multiuser environment. If the functions are written carefully, the system can be reliable. And even if some of the ACID properties are compromised due to lack of real RM, the system is still useful for its undo/redo capability.What are the drawbacks?
The reliability of the system rests on the reliability of each involved function. One buggy function can break the transaction.What about non-undoable actions?
Non-undoable actions (like sending an email, permanently deleting files) should be executed outside the scope of transaction.HOMEPAGE
Please visit the project's homepage at <https://metacpan.org/release/Rinci>.SOURCE
Source repository is at <https://github.com/sharyanto/perl-Rinci>.BUGS
Please report any bugs or feature requests on the bugtracker website <https://rt.cpan.org/Public/Dist/Display.html?Name=Rinci>When submitting a bug or request, please include a test-file or a patch to an existing test-file that illustrates the bug or desired feature.
AUTHOR
perlancar <[email protected]>COPYRIGHT AND LICENSE
This software is copyright (c) 2016 by [email protected].This is free software; you can redistribute it and/or modify it under the same terms as the Perl 5 programming language system itself.