Deleting the files should be good.

One tricky thing are also dependencies among files. We need to make sure that file "testrealm.json" with realm definition is imported before users file "testrealm-users-X.json" . Also master realm should be always imported first. We can solve this by divide files into groups:
1) File with master realm (if exists)
2) Other files with realm definitions
3) All user's files
We need to ensure that import of each group will start after previous group is fully finished. But this is solvable too.


So far I can see this as my favourite workflow:

a) When import is triggered, coordinator will ask all cluster nodes for file checksums they have. Each node will check DB and delete the files, which are already in DB (file with that checksum was already imported in previous import iteration). Then node send the checksums (just those which were not deleted) and to which group each checksum belongs (master, realms, users)

b) Import of group1 starts. Coordinator will elect the node, which has the file with particular checksum available. For example if checksum "abcd" is available on node1 and node3, any of those 2 nodes are elected.

c) Once import of file "abcd" is successful, node will save the record "abcd - IMPORT SUCCESS" to DB and send message to coordinator about finished import.

d) If node1 crashes, the import of "abcd" file is immediatelly re-triggered by coordinator on node3. Similarly if file with checksum "abcd" is not available on node1 (this can happen if admin edited the file on node1 and saved it. The checksum of the file will be changed in this case), the import is re-triggered on node3 too.

e) Once group1 (master realm) is finished, go back to step (b) and start with group2 (other realms), then again with group3 (user's files)

f) The checksums are collected just at startup (step a). If someone change the file on any node during import iteration, it will be new checksum and this one will be ignored until this import iteration is finished.

g) If new node joins the cluster, it can start helping with import as long as he has the files with checksums collected at step (a).

At the end of the import iteration, there are DB records of all successfully imported checksums. When new import iteration is triggered, the nodes will delete the files already imported (we can also do that at the end of the import iteration too)

Maybe it sounds a bit complicated, but I can see this as most performant workflow and vulnerable to crashes of any cluster node. Also it's vulnerable to the situation when admin changed some file when import is in progress.

Marek


On 11/11/15 16:01, Stian Thorgersen wrote:
What about if we remove files as they are being imported?

Something like:

* When we detect a new file on a node, we send the name + checksum to the cordinator
* The cordinator then checks if this file has already been imported
* If it's imported it sends out a message stating that file with name + checksum is already imported and all nodes delete this file
* If it's not imported it picks one node that is responsible to import the file (that would be the first node that sends the message about the file+checksum). This node will rename the file to .importing
* Once the file has been imported it's renamed to .imported or .failed

On 11 November 2015 at 15:53, Marek Posolda <mposolda@redhat.com> wrote:
On 11/11/15 15:51, Marek Posolda wrote:
On 11/11/15 15:36, Stian Thorgersen wrote:


On 11 November 2015 at 15:23, Marek Posolda <mposolda@redhat.com> wrote:
On 11/11/15 09:01, Stian Thorgersen wrote:


On 10 November 2015 at 16:11, Marek Posolda <mposolda@redhat.com> wrote:
On 09/11/15 14:09, Stian Thorgersen wrote:


On 9 November 2015 at 13:35, Sebastien Blanc <sblanc@redhat.com> wrote:
That would be really nice indeed !
But are the markers files not enough, instead of also having a table in the DB ?

We need a way to prevent multiple nodes in a cluster to import the same file. For example on Kerberos you end up spinning up multiple instances of the same Docker image.
I bet you meant 'Kubernetes' :-)

Yup
 


+1 for the improvements. Besides those I think that earlier or later, we will need to solve long-running export+import where you want to import 100.000 users.

+1
 

As I mentioned in another mail few weeks ago, we can have:

1) Table with the progress (51.000 users already imported, around 49.000 remaining etc.)

We would still need to split into multiple files in either case. Having a single json file with 100K users is probably not going to perform very well. So what I proposed would actually work for long-running import as well. If each file has a manageable amount of users (say ~5 min to import) then each file will be marked as imported or failed. At least for now I don't think we should do smaller batches than one file. As long as one file is imported within the same TX then it's an all or nothing import.
 
2) Concurrency and dividing the work among cluster nodes (Node1 will import 50.000 users and node2 another 50.000 users)

This would be solved as well. Each node picks up a file that's not processed yet. Marks it in the DB and then gets to process it.
 
3) Failover (Import won't be completely broken if cluster node crashes after import 90.000, but can continue on other cluster nodes)

I think the stuff I did recently for pre-loading offline sessions at startup could be reused for this stuff too and it can handle (2) and (3) . Also it can handle parallel import triggered from more cluster nodes.

For example: currently if you trigger kubernetes with 2 cluster nodes, both nodes will start to import same file at the same time because import triggered by node1 is not yet finished before node2 is started, so there is not yet existing DB record that file is already imported. With the stuff I did, just the coordinator (node1) will start the import . Node2 will wait until import triggered by node1 is finished, but at the same time it can "help" to import some users (pages) if coordinator asks him to do so. This impl is based on infinispan distributed executor service http://infinispan.org/docs/5.3.x/user_guide/user_guide.html#_distributed_execution_framework .

The DB record needs to be created before a node tries to import it, including a timestamp when it started the import. It should then be updated once the import is completed, with the result. Using the distributed execution framework sounds like a good idea though. How do you prevent scheduling the same job multiple times? For example if all nodes on startup scan the import folder and simply import everything they find, then there will be multiple of the same job. Not really a big deal as the first thing the job should do is check if there's a record in the DB already.
 With distributed executor, it's the cluster coordinator, which coordinates which node would import what. It will send messages to cluster nodes like "Hey, please import the file testrealm-users-3.json with timestamp abcd123" .

After node finishes the job, it notifies coordinator and coordinator will insert DB record and mark it as finished. So there is no DB record inserted before node starts import, because whole coordination is handled by the coordinator. Also there will never be same file imported more times by different cluster nodes.

Only exception would be if cluster node crashes before import is finished. Then it needs to be reimported by other cluster node, but that's the case with DB locks as well.

IMO the DB locks approach doesn't handle well crash of some cluster node. For example when node2 crashes unexpectedly when it's importing the file testrealm-users-3.json, the DB lock is held by this node, so other cluster nodes can't start on importing the file (until timeout occurs.)

On the other hand, distributed executor approach may have issues if there is inconsistent content of the standalone/import directory among cluster nodes. However it can be solved, so that each node will need to send checksums of the files it has and coordinator will need to ensure that file with checksum "abcd123" is assigned just to the node which has this file.

With Docker/Kubernetes all nodes would have the same files. At least initially. Would be nice if we could come up with a solution where you can just drop an additional file onto any node and have it imported.
Exactly, was thinking about Docker too. Here we don't have any issue at all.

The main question here is, do we want to support the scenario when various cluster nodes have different content? As I mentioned, distributed coordinator can handle it, so that each cluster node will send the checksums of the files it has and coordinator will always assign to node just the checksums, which it has.

However regardless of distributed executor approach or DB locks approach, there may be still the issues. For example:
1) The file testrealm.json with checksum "abc" is triggered for import on node1
2) At the same time, admin will do some minor change in this file on node2 and save it. This will mean that checksum of the file on node2 will be changed to "def"
3) Node2 will trigger import of that file. So we have both node1 and node2 importing same file concurrently because the previously retrieved lock was for "abc" checksum, but now checksum is "def"

This problem will be with both DB lock and DistributedExecutor approaches though...
Possible solution for this issue is, that when import is already in progress, the newly added or changed checksums will be ignored. The checksums will be always checked just at start of the import.

Marek


Marek
 


Marek

 


Marek


 
 

On Mon, Nov 9, 2015 at 1:20 PM, Stian Thorgersen <sthorger@redhat.com> wrote:
Currently we support importing a complete realm definition using the import/export feature. Issues with the current approach is:

* Only complete realm - not possible to add to an existing realm
* No good feedback if import was successful or not
* Use of system properties to initiate the import is not very user friendly
* Not very elegant for provisioning. For example a Docker image that want's to bundle some initial setup ends up always running the import of a realm, which is skipped if realm exists

To solve this I've come up with the following proposal:

Allow dropping representations to be imported into 'standalone/import'. This should support creating a new realm as well as importing into an existing realm. When importing into an existing realm we will have an import strategy that is used to configure what happens if a resource exists (user, role, identity provider, user federtation provider). The import strategies are:

* Skip - existing resources are skipped,
* Fail - if any resource exists nothing is imported
* Overwrite - any existing resources are deleted.

The directory will be scanned at startup, but there will also be an option to monitor this directory at runtime.

To prevent a file being imported multiple times (also to make sure only one node in a cluster imports) we will have a table in the database that contains what files was imported, from what node, date and result (including a list of what resources where imported, which was not, and stack trace if applicable). The primary key will be the checksum of the file. We will also add marker files (<json file>.imported or <json file>.failed). The contents of the marker files will be a json object with date imported, outcome (including stack trace if applicable) as well as a complete list of what resources was successfully imported, what where not.

The files will also allow resolving system properties and environment variables. For example:

{
    "secret": "${env.MYCLIENT_SECRET}"
}

This will be very convenient for example with Docker as it would be very easy to create a Docker image that extends ours to add a few clients and users.

It will also be convenient for examples as it will make it possible to add the required clients and users to an existing realm.



_______________________________________________
keycloak-dev mailing list
keycloak-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/keycloak-dev




_______________________________________________
keycloak-dev mailing list
keycloak-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/keycloak-dev







_______________________________________________
keycloak-dev mailing list
keycloak-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/keycloak-dev