[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15552593#comment-15552593 ] Varun Saxena edited comment on YARN-5585 at 10/6/16 5:30 PM: - bq. Given entities are sorted in ascending order, at some extent latest fist order can be achieve by doing reverse scan. I had tried this for yarn-containers and works fine. Reverse scan would work fine but how do we decide which entity type would need it and which won't. By the way we need container IDs' in the reverse order too ? IIRC, in one of the calls Li mentioned lexicographic order should be fine for new Web UI. If required we can have special handling for YARN specific entities like App attempts and Containers, just like we have for apps. No matter what we do, it should be consistent across all entities. We can also have another query param to indicate reverse lexicographic order is required. bq. IIUC, AM can delegate collector address to any of its running containers to publish its own data. TimelineClient can not be restricted to only AM. True. In a secure setup, AM can even pass on the token. The point is we support talking to AM only. AM can then delegate its work to anyone. But the concern here was that prefix will have to be passed around by AM via a new protocol. So if application wants to support delegating work to other processes, it anyways needs to open new protocol. So I guess this concern is not specific to prefix. Correct ? However, would be useful if you can tell the use case of multiple JVMs'. Same DAGs' can be executed by different processes. This would help us understanding the use case and decide how best to support it. was (Author: varun_saxena): bq. Given entities are sorted in ascending order, at some extent latest fist order can be achieve by doing reverse scan. I had tried this for yarn-containers and works fine. Reverse scan would work fine but how do we decide which entity type would need it and which won't. By the way we need container IDs' in the reverse order too ? IIRC, in one of the calls Li mentioned lexicographic order should be fine for new Web UI. If required we can have special handling for YARN specific entities like App attempts and Containers, just like we have for apps. No matter what we do, it should be consistent across all entities. We can also have another query param to indicate reverse lexicographic order is required. bq. IIUC, AM can delegate collector address to any of its running containers to publish its own data. TimelineClient can not be restricted to only AM. True. In a secure setup, AM can even pass on the token. The point is we support talking to AM only. AM can then delegate its work to anyone. But the concern here was that prefix will have to be passed around by AM via a new protocol. So if application wants to support delegating work to other processes, it anyways needs to open new protocol. So I guess this concern is not specific to prefix. Correct ? However, would be useful if you can tell the use case of multiple JVMs'. Same DAGs' can be executed by different processes. This would help us thing > [Atsv2] Add a new filter fromId in REST endpoints > - > > Key: YARN-5585 > URL: https://issues.apache.org/jira/browse/YARN-5585 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelinereader >Reporter: Rohith Sharma K S >Assignee: Rohith Sharma K S >Priority: Critical > Attachments: 0001-YARN-5585.patch, YARN-5585-workaround.patch, > YARN-5585.v0.patch > > > TimelineReader REST API's provides lot of filters to retrieve the > applications. Along with those, it would be good to add new filter i.e fromId > so that entities can be retrieved after the fromId. > Current Behavior : Default limit is set to 100. If there are 1000 entities > then REST call gives first/last 100 entities. How to retrieve next set of 100 > entities i.e 101 to 200 OR 900 to 801? > Example : If applications are stored database, app-1 app-2 ... app-10. > *getApps?limit=5* gives app-1 to app-5. But to retrieve next 5 apps, there is > no way to achieve this. > So proposal is to have fromId in the filter like > *getApps?limit=5&=app-5* which gives list of apps from app-6 to > app-10. > Since ATS is targeting large number of entities storage, it is very common > use case to get next set of entities using fromId rather than querying all > the entites. This is very useful for pagination in web UI. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15552144#comment-15552144 ] Varun Saxena edited comment on YARN-5585 at 10/6/16 3:01 PM: - bq. I was thinking to use same REST API for both by using SingleColumnFilter. One cons I see is table scan for all the entityType i.e reflect in read performance. We should not use SingleColumnValueFilter if we know the prefix because as you said former will lead to a relatively slower read performance. Basically we need to differentiate between not having a prefix for the entity type and user unable to supply it. bq. I would have thought that we store the entities in the reverse entity id order, but it appears that the entity id is encoded into the row key as is (EntityRowKey). Am I reading that right? If so, this is a bug to fix. Entity IDs' can be anything. Even a completely alphabetical sequence can be an entity ID. So it will not be possible to define a reverse order for every generic entity ID. Is this your question ? bq. Firstly about multi JVM which makes application programmer to define new protocol for transferring prefixId. Trying to understand this more. Can same DAG be executed by multiple Tez AMs' ? bq. Secondly, what if users misses providing an prefixId in subsequent updates.? This should be caught during integration phase. Right ? was (Author: varun_saxena): bq. I was thinking to use same REST API for both by using SingleColumnFilter. One cons I see is table scan for all the entityType i.e reflect in read performance. We should not use SingleColumnValueFilter if we know the prefix because as you said former will lead to a relatively slower read performance. Basically we need to differentiate between having a prefix for the entity type and user unable to supply it. bq. I would have thought that we store the entities in the reverse entity id order, but it appears that the entity id is encoded into the row key as is (EntityRowKey). Am I reading that right? If so, this is a bug to fix. Entity IDs' can be anything. Even a completely alphabetical sequence can be an entity ID. So it will not be possible to define a reverse order for every generic entity ID. Is this your question ? bq. Firstly about multi JVM which makes application programmer to define new protocol for transferring prefixId. Trying to understand this more. Can same DAG be executed by multiple Tez AMs' ? bq. Secondly, what if users misses providing an prefixId in subsequent updates.? This should be caught during integration phase. Right ? > [Atsv2] Add a new filter fromId in REST endpoints > - > > Key: YARN-5585 > URL: https://issues.apache.org/jira/browse/YARN-5585 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelinereader >Reporter: Rohith Sharma K S >Assignee: Rohith Sharma K S >Priority: Critical > Attachments: 0001-YARN-5585.patch, YARN-5585-workaround.patch, > YARN-5585.v0.patch > > > TimelineReader REST API's provides lot of filters to retrieve the > applications. Along with those, it would be good to add new filter i.e fromId > so that entities can be retrieved after the fromId. > Current Behavior : Default limit is set to 100. If there are 1000 entities > then REST call gives first/last 100 entities. How to retrieve next set of 100 > entities i.e 101 to 200 OR 900 to 801? > Example : If applications are stored database, app-1 app-2 ... app-10. > *getApps?limit=5* gives app-1 to app-5. But to retrieve next 5 apps, there is > no way to achieve this. > So proposal is to have fromId in the filter like > *getApps?limit=5&=app-5* which gives list of apps from app-6 to > app-10. > Since ATS is targeting large number of entities storage, it is very common > use case to get next set of entities using fromId rather than querying all > the entites. This is very useful for pagination in web UI. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15545818#comment-15545818 ] Varun Saxena edited comment on YARN-5585 at 10/4/16 4:06 PM: - Thanks [~rohithsharma] for the patch. Few comments. # Intention behind having ID_PREFIX in EntityColumn ? According to me, we need not store prefix in the column. Is it because we want to read it back and send it to client ? # No need of GenericEntityReader#calculateTheClosestNextRowKeyForPrefix. Scan#setRowPrefixFilter will do it for you. We should call it the same way as was done previously. # As entity ID prefix is a long, EntityRowKeyConverter#SEGMENT_SIZES should have new segment as Bytes.SIZEOF_LONG. It is currently given as VARIABLE_SIZE. Same change in TestRowKeys. # In EntityRowKeyConverter#encode, no need to invert entity id prefix. We will take prefix as-is. Sender can publish the entity with inverted prefix if he wants contents in descending order (say). We can probably add something to TimelineUtils to invert it, if required, which then clients can use. # In GenericEntityReader#parseEntity we should fetch id prefix from result set and setIdPrefix in TimelineEntity to be returned back. This will be useful for clients when they want to set fromPrefix (will be useful in Tez UI use case). # Javadoc in TimelineReader should be changed. It currently says entities would be sorted by created time which is no longer true. {code} * @return A set of TimelineEntity instances of the given entity *type in the given context scope which matches the given predicates *ordered by created time, descending. Each entity will only contain the *metadata(id, type and created time) plus the given fields to retrieve. {code} # We should also update documentation to reflect id prefix. was (Author: varun_saxena): Thanks [~rohithsharma] for the patch. Few comments. # Intention behind having ID_PREFIX in EntityColumn ? According to me, we need not store prefix in the column. Is it because we want to read it back and send it to client ? # No need of GenericEntityReader#calculateTheClosestNextRowKeyForPrefix. Scan#setRowPrefixFilter will do it for you. We should call it the same way as was done previously. # As entity ID prefix is a long, EntityRowKeyConverter#SEGMENT_SIZES should have new segment as Bytes.SIZEOF_LONG. Same change in TestRowKeys. # In EntityRowKeyConverter#encode, no need to invert entity id prefix. We will take prefix as-is. Sender can publish the entity with inverted prefix if he wants contents in descending order (say). We can probably add something to TimelineUtils to invert it, if required, which then clients can use. # In GenericEntityReader#parseEntity we should fetch id prefix from result set and setIdPrefix in TimelineEntity to be returned back. This will be useful for clients when they want to set fromPrefix (will be useful in Tez UI use case). # Javadoc in TimelineReader should be changed. It currently says entities would be sorted by created time which is no longer true. {code} * @return A set of TimelineEntity instances of the given entity *type in the given context scope which matches the given predicates *ordered by created time, descending. Each entity will only contain the *metadata(id, type and created time) plus the given fields to retrieve. {code} # We should also update documentation to reflect id prefix. > [Atsv2] Add a new filter fromId in REST endpoints > - > > Key: YARN-5585 > URL: https://issues.apache.org/jira/browse/YARN-5585 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelinereader >Reporter: Rohith Sharma K S >Assignee: Rohith Sharma K S >Priority: Critical > Attachments: 0001-YARN-5585.patch, YARN-5585-workaround.patch, > YARN-5585.v0.patch > > > TimelineReader REST API's provides lot of filters to retrieve the > applications. Along with those, it would be good to add new filter i.e fromId > so that entities can be retrieved after the fromId. > Current Behavior : Default limit is set to 100. If there are 1000 entities > then REST call gives first/last 100 entities. How to retrieve next set of 100 > entities i.e 101 to 200 OR 900 to 801? > Example : If applications are stored database, app-1 app-2 ... app-10. > *getApps?limit=5* gives app-1 to app-5. But to retrieve next 5 apps, there is > no way to achieve this. > So proposal is to have fromId in the filter like > *getApps?limit=5&=app-5* which gives list of apps from app-6 to > app-10. > Since ATS is targeting large number of entities storage, it is very common > use case to get next set of entities using fromId rather than querying all > the entites. This is very useful for pagination in
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15533830#comment-15533830 ] Vrushali C edited comment on YARN-5585 at 9/29/16 7:40 PM: --- Thanks [~rohithsharma] for the summary. bq. 2. By default, use createdTime as entityPrefixId. Also, that means, frameworks which don't want to use the entity id prefix have to explicitly specify a null prefix (or a special value that means null). All the same, it will be really good to mention in the docs for clients that they should do the following. {code:title=TimelineWriterClient.java} entity.setEntityPrefix(createdTime); client.writeEntity(entity); // pseudo-code {code} bq. For the REST end point, we can support fromEntityPrefixId will become combination of entityPrefixId+entityId which can be used for pagination I think pagination handling should be more generic than depending on something like "fromEntityPrefixId". REST queries should simply ask for top N records with the understanding that the records are returned in sorted order of entity prefixes. For the next page of results, the client sends back the last row returned's key/entity prefix. For a rest query, if the "startFrom" query param is present, the scan starts from "startFrom" prefix value and returns the next N such records. was (Author: vrushalic): Thanks [~rohithsharma] for the summary. bq. 2. By default, use createdTime as entityPrefixId. Also, that means, frameworks which don't want to use the entity id prefix have to explicitly specify a null prefix (or a special value that means null). All the same, it will be really good to mention in the docs for clients that they should do the following. {code:title=TimelineWriterClient.java} entity.setEntityPrefix(createdTime); client.writeEntity(entity); // pseudo-code {code} bq. For the REST end point, we can support fromEntityPrefixId will become combination of entityPrefixId+entityId which can be used for pagination I think pagination handling should be more generic than depending on something like "fromEntityPrefixId". REST queries should simply ask for top N records with the understanding that the records are returned in sorted order of entity prefixes. For the next page of results, the client sends back the last row returned's key/entity prefix. For a rest query, if the "startFrom" query param is present, the scan starts from "startFrom" prefix value and returns the next N such records. > [Atsv2] Add a new filter fromId in REST endpoints > - > > Key: YARN-5585 > URL: https://issues.apache.org/jira/browse/YARN-5585 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelinereader >Reporter: Rohith Sharma K S >Assignee: Rohith Sharma K S >Priority: Critical > Attachments: YARN-5585-workaround.patch, YARN-5585.v0.patch > > > TimelineReader REST API's provides lot of filters to retrieve the > applications. Along with those, it would be good to add new filter i.e fromId > so that entities can be retrieved after the fromId. > Current Behavior : Default limit is set to 100. If there are 1000 entities > then REST call gives first/last 100 entities. How to retrieve next set of 100 > entities i.e 101 to 200 OR 900 to 801? > Example : If applications are stored database, app-1 app-2 ... app-10. > *getApps?limit=5* gives app-1 to app-5. But to retrieve next 5 apps, there is > no way to achieve this. > So proposal is to have fromId in the filter like > *getApps?limit=5&=app-5* which gives list of apps from app-6 to > app-10. > Since ATS is targeting large number of entities storage, it is very common > use case to get next set of entities using fromId rather than querying all > the entites. This is very useful for pagination in web UI. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15523177#comment-15523177 ] Varun Saxena edited comment on YARN-5585 at 9/26/16 2:13 PM: - bq. When scan is performed on rows, the ResultScanner gives in the order of lexicographical order only. I could not get where does this entityIdPrefix will be used? Is it from storage or readerservere? Entity ID prefix will be supplied by Tez in your case, while publishing entities, and can be inverse of created time if you want rows to be sorted in a descending order by created time. TimelineEntity class will now carry a prefix too. bq. Does new tables separate or same-existing? Same tables. Just the row key changes if you are not happy with lexicographic order. was (Author: varun_saxena): bq. When scan is performed on rows, the ResultScanner gives in the order of lexicographical order only. I could not get where does this entityIdPrefix will be used? Is it from storage or readerservere? Entity ID prefix will be supplied by Tez in your case and can be inverse of created time if you want rows to be sorted in a descending order by created time. TimelineEntity class will now carry a prefix too. bq. Does new tables separate or same-existing? Same tables. Just the row key changes if you are not happy with lexicographic order. > [Atsv2] Add a new filter fromId in REST endpoints > - > > Key: YARN-5585 > URL: https://issues.apache.org/jira/browse/YARN-5585 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelinereader >Reporter: Rohith Sharma K S >Assignee: Rohith Sharma K S >Priority: Critical > Attachments: YARN-5585-workaround.patch, YARN-5585.v0.patch > > > TimelineReader REST API's provides lot of filters to retrieve the > applications. Along with those, it would be good to add new filter i.e fromId > so that entities can be retrieved after the fromId. > Current Behavior : Default limit is set to 100. If there are 1000 entities > then REST call gives first/last 100 entities. How to retrieve next set of 100 > entities i.e 101 to 200 OR 900 to 801? > Example : If applications are stored database, app-1 app-2 ... app-10. > *getApps?limit=5* gives app-1 to app-5. But to retrieve next 5 apps, there is > no way to achieve this. > So proposal is to have fromId in the filter like > *getApps?limit=5&=app-5* which gives list of apps from app-6 to > app-10. > Since ATS is targeting large number of entities storage, it is very common > use case to get next set of entities using fromId rather than querying all > the entites. This is very useful for pagination in web UI. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15514243#comment-15514243 ] Varun Saxena edited comment on YARN-5585 at 9/22/16 7:25 PM: - Summarizing the solution we decided upon in the call. * We will now return entities from entity table in a lexicographic order of entity IDs' * To achieve a different sort order, we will provide a mechanism for applications to provide an entity ID prefix which can be set in the TimelineEntity object while publishing the entity. * This entityId prefix will be part of the row key in entity table. As the name suggests, it will be present just before the entity ID. Applications can choose to provide no entity ID prefix if they are happy with the lexicographic sort order. So the row key now will be {{cluster!user!flow!flowrun!app!entitytype!\{entityidprefix\}!\{entityid\}}} * Entity ID will also be stored under a column qualifier too (being done already). * Entity ID prefix can be a number (say long) as numbers generally provide a natural sort ordering. However, this needs to be finalized. Keep it as a string ? * When querying multiple entities, we will return the top N entities decided by limit in a lexicographic order of entity ID prefix + entity ID (i.e. if entity ID prefix is supplied). fromID filter can now be something like fromIDPrefix (say) or a similar filter which provides prefix + ID to support pagination. * While querying a single entity, prefix can be supplied as a query param. If supplied, it will be a Get, otherwise we need to have a Scan with SingleColumnValueFilter on entity ID (this will be comparatively slower). We can have a separate REST endpoint to distinguish between prefix based queries and non prefix based queries. We need to distinguish between the case where for an entity prefix has not been specified on the write path and prefix not just supplied at the read path (even if it was supplied at the write path). This needs to be finalized. * Prefix will also be returned as part of TimelineEntity object in response. cc [~jrottinghuis], [~sjlee0], [~vrushalic], [~gtCarrera9]. Hope this covers everything. The reason this solution was chosen was that we thought in UI use cases a single entity read would typically be followed listing of multiple entities and hence prefix would be known. This does not mean however, that we will not provide a mechanism to fetch entity if prefix wasn't given. We can use a single column value filter then. Moreover, this solution overall had lesser write or read penalty compared to solutions listed above. was (Author: varun_saxena): Summarizing the solution we decided upon in the call. * We will now return entities from entity table in a lexicographic order of entity IDs' * To achieve a different sort order, we will provide a mechanism for applications to provide an entity ID prefix which can be set in the TimelineEntity object while writing the entity to backend. * This entityId prefix will be part of the row key in entity table. As the name suggests, it will be present just before the entity ID. Applications can choose to provide no entity ID prefix if they are happy with the lexicographic sort order. So the row key now will be {{cluster!user!flow!flowrun!app!entitytype!\{entityidprefix\}!\{entityid\}}} * Entity ID will also be stored under a column qualifier too (being done already). * Entity ID prefix can be a number (say long) as numbers generally provide a natural sort ordering. However, this needs to be finalized. Keep it as a string ? * When querying multiple entities, we will return the top N entities decided by limit in a lexicographic order of entity ID prefix + entity ID (i.e. if entity ID prefix is supplied). fromID filter can now be something like fromIDPrefix (say) or a similar filter which provides prefix + ID to support pagination. * While querying a single entity, prefix can be supplied as a query param. If supplied, it will be a Get, otherwise we need to have a Scan with SingleColumnValueFilter on entity ID (this will be comparatively slower). We can have a separate REST endpoint to distinguish between prefix based queries and non prefix based queries. We need to distinguish between the case where for an entity prefix has not been specified on the write path and prefix not just supplied at the read path (even if it was supplied at the write path). This needs to be finalized. * Prefix will also be returned as part of TimelineEntity object in response. cc [~jrottinghuis], [~sjlee0], [~vrushalic], [~gtCarrera9]. Hope this covers everything. The reason this solution was chosen was that we thought in UI use cases a single entity read would typically be followed listing of multiple entities and hence prefix would be known. This does not mean however, that we will not provide a mechanism to fetch entity if prefix
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15512200#comment-15512200 ] Sangjin Lee edited comment on YARN-5585 at 9/22/16 5:26 AM: I am also catching up on this discussion (sorry it got delayed). Generally I am in agreement with Varun and Vrushali on possible approaches. I'd like to add a few more thoughts to refine the idea. (1) supporting chronological order sorting I think that even for framework-specific entities (e.g. tez vertices, MR task entities, etc.), the "sorting" order cannot be completely arbitrary. Because we have a strong design decision on reflecting recency in the row keys, the natural sorting order should be the *chronological order*, or strange things would result. For YARN entities, the id order would satisfy this for the most part (and ditto for MR entities). If tez can craft the id's such that the lexicographical order is also the chronological order, that would be by far the simplest solution to the problem. I'm not sure how feasible it is for tez to add padding etc. to preserve the chronological order in the entity id's. [~rohithsharma], can we change the id's to order them properly? If the framework cannot make the id lexicographical order the same as the chronological order, then we might have to introduce the notion of bytes provided by the framework (and an auxiliary table) to support this as suggested by Vrushali and Varun. But that would be at the some cost. All things being equal, I would love not to populate another table on the write path. Also note that we still need to support single-entity queries in this case (i.e. queries by entity id). How would we be able to support queries by id in this case? (2) setting the created time field In timeline service v.2, the strong assumption/requirement is that the created time is set by the client. It sounds like the current tez code does not set the created time. It should be set. That's the contract we're using. We're not really expecting an empty created time when we write them. (3) TimelineEntity.compareTo() It is a good catch by Rohith. It escaped the review, but it does appear that the id sorting if created time is empty is the opposite of what it should be. The string should be sorted by the descending order, but the current code is doing the opposite. This should be fixed. We can either fix it here or can open a separate subtask to fix it. Either way, we should fix it. was (Author: sjlee0): I am also catching up on this discussion (sorry it got delayed). Generally I am in agreement with Varun and Vrushali on possible approaches. I'd like to add a few more thoughts to refine the idea. (1) supporting chronological order sorting I think that even for framework-specific entities (e.g. tez vertices, MR task entities, etc.), the "sorting" order cannot be completely arbitrary. Because we have a strong design decision on reflecting recency in the row keys, the natural sorting order should be the *chronological order*, or strange things would result. For YARN entities, the id order would satisfy this for the most part (and ditto for MR entities). If tez can craft the id's such that the lexicographical order is also the chronological order, that would be by far the simplest solution to the problem. I'm not sure how feasible it is for tez to add padding etc. to preserve the chronological order in the entity id's. [~rohithsharma], can we change the id's to order them properly? If the framework cannot make the id lexicographical order the same as the chronological order, then we might have to introduce the notion of bytes provided by the framework (and an auxiliary table) to support this as suggested by Vrushali and Varun. But that would be at the some cost. All things being equal, I would love not to populate another table on the write path. Also note that we still need to be able to support single-entity queries in this case (i.e. queries by entity id). How would we able to support queries by id in this case? (2) setting the created time field In timeline service v.2, the strong assumption/requirement is that the created time is set by the client. It sounds like the current tez code does not set the created time. I think it should be set. That's the contract we're using. We're not really expecting an empty created time when we write them. (3) TimelineEntity.compareTo() It is a good catch by Rohith. It escaped the review, but it does appear that the id sorting if created time is empty is the opposite of what it should be. The string should be sorted by the descending order, but the current code is doing the opposite. This should be fixed. We can either fix it here or can open a separate subtask to fix it. Either way, we should fix it. > [Atsv2] Add a new filter fromId in REST endpoints >
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15512079#comment-15512079 ] Vrushali C edited comment on YARN-5585 at 9/22/16 4:23 AM: --- I have been thinking more on this. I think if there is a concern about having the same entity data in two tables, what we could do is, set a TTL (time to live) on the cells in the auxiliary table. That way, for some period of time we store data in two places but then it gets cleaned up. For example, if the Tez UI queries for data in the auxiliary table for a job that ran 1 year back, then say, it does not exist anymore in the auxiliary table since it got cleaned up by hbase. Now the Tez UI can try querying the regular table. Or the auxiliary REST api call can take a parameter that says if data is not found in auxiliary table, please query the regular entity table and the rest call would perhaps then take a little longer to return. Since we are querying for something that ran 1 year back, I believe we can wait for an extra moment for the call to return. This way, we store data in two tables for a brief time period, rely on hbase to clean up cells as per their TTL and provide a way for frameworks to store/query their data in harmony with timeline service storage. was (Author: vrushalic): I have been thinking more on this. I think if there is a concern about having the same entity data in two tables, what we could do is, set a TTL (time to live) on the cells in the auxiliary table. That way, for some period of time we store data in two places but then it gets cleaned up. For example, if Tez UI queries for data in the auxiliary table for a job that ran 1 year back, then say, it does not exist anymore in the auxiliary table since it got cleaned up by hbase. Now the Tez UI can try querying the regular table. Or the auxiliary REST api call can take a parameter that says if data is not found in auxiliary table, please query the regular entity table and the rest call would perhaps then take a little longer to return. Since we are querying for something that ran 1 year back, I believe we can wait for an extra moment for the call to return. This way, we store data in two tables for a brief time period, rely on hbase to clean up cells as per their TTL and provide a way for frameworks to store/query their data in harmony with timeline service storage. > [Atsv2] Add a new filter fromId in REST endpoints > - > > Key: YARN-5585 > URL: https://issues.apache.org/jira/browse/YARN-5585 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelinereader >Reporter: Rohith Sharma K S >Assignee: Rohith Sharma K S >Priority: Critical > Attachments: YARN-5585.v0.patch > > > TimelineReader REST API's provides lot of filters to retrieve the > applications. Along with those, it would be good to add new filter i.e fromId > so that entities can be retrieved after the fromId. > Current Behavior : Default limit is set to 100. If there are 1000 entities > then REST call gives first/last 100 entities. How to retrieve next set of 100 > entities i.e 101 to 200 OR 900 to 801? > Example : If applications are stored database, app-1 app-2 ... app-10. > *getApps?limit=5* gives app-1 to app-5. But to retrieve next 5 apps, there is > no way to achieve this. > So proposal is to have fromId in the filter like > *getApps?limit=5&=app-5* which gives list of apps from app-6 to > app-10. > Since ATS is targeting large number of entities storage, it is very common > use case to get next set of entities using fromId rather than querying all > the entites. This is very useful for pagination in web UI. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15512079#comment-15512079 ] Vrushali C edited comment on YARN-5585 at 9/22/16 4:22 AM: --- I have been thinking more on this. I think if there is a concern about having the same entity data in two tables, what we could do is, set a TTL (time to live) on the cells in the auxiliary table. That way, for some period of time we store data in two places but then it gets cleaned up. For example, if Tez UI queries for data in the auxiliary table for a job that ran 1 year back, then say, it does not exist anymore in the auxiliary table since it got cleaned up by hbase. Now the Tez UI can try querying the regular table. Or the auxiliary REST api call can take a parameter that says if data is not found in auxiliary table, please query the regular entity table and the rest call would perhaps then take a little longer to return. Since we are querying for something that ran 1 year back, I believe we can wait for an extra moment for the call to return. This way, we store data in two tables for a brief time period, rely on hbase to clean up cells as per their TTL and provide a way for frameworks to store/query their data in harmony with timeline service storage. was (Author: vrushalic): I have been thinking more on this. I think if there is a concern about having the same entity data in two tables, what we could do is, set a TTL (time to live) on the cells in the auxiliary table. That way, for some period of time time we store data in two places but then it gets cleaned up. For example, if Tez UI queries for data in the auxiliary table for a job that ran 1 year back, then say, it does not exist anymore in the auxiliary table since it got cleaned up by hbase. Now the Tez UI can try querying the regular table. Or the auxiliary REST api call can take a parameter that says if data is not found in auxiliary table, please query the regular entity table and the rest call would perhaps then take a little longer to return. Since we are querying for something that ran 1 year back, I believe we can wait for an extra moment for the call to return. This way, we store data in two tables for a brief time period, rely on hbase to clean up cells as per their TTL and provide a way for frameworks to store/query their data in harmony with timeline service storage. > [Atsv2] Add a new filter fromId in REST endpoints > - > > Key: YARN-5585 > URL: https://issues.apache.org/jira/browse/YARN-5585 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelinereader >Reporter: Rohith Sharma K S >Assignee: Rohith Sharma K S >Priority: Critical > Attachments: YARN-5585.v0.patch > > > TimelineReader REST API's provides lot of filters to retrieve the > applications. Along with those, it would be good to add new filter i.e fromId > so that entities can be retrieved after the fromId. > Current Behavior : Default limit is set to 100. If there are 1000 entities > then REST call gives first/last 100 entities. How to retrieve next set of 100 > entities i.e 101 to 200 OR 900 to 801? > Example : If applications are stored database, app-1 app-2 ... app-10. > *getApps?limit=5* gives app-1 to app-5. But to retrieve next 5 apps, there is > no way to achieve this. > So proposal is to have fromId in the filter like > *getApps?limit=5&=app-5* which gives list of apps from app-6 to > app-10. > Since ATS is targeting large number of entities storage, it is very common > use case to get next set of entities using fromId rather than querying all > the entites. This is very useful for pagination in web UI. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15511129#comment-15511129 ] Vrushali C edited comment on YARN-5585 at 9/21/16 8:51 PM: --- Catching up on this thread. I have tried to read through all the comments and discussions on this jira but please correct me if I am mistaken. Two objectives here: 1) We are looking for a way to paginate results/response. 2) Ability to return sorted results in the rest response (sorted on something other than row key) Thoughts on these: - We are looking for a way to paginate results/response. This pagination requirement is independent of any particular framework like Tez. With hRaven, our experience has been that more often than not, we end up enabling pagination support for most APIs. So, in general, our rest api calls should support pagination. Pagination via the REST query: This involves, in a generic fashion, being able to send in a “startFromRowKey” in the rest query. Say we extend our rest apis to accept such a parameter, it becomes generic enough to fetch N rows after this particular “startFromRowKey” value. The first rest api call will not send in anything, but each rest response will return “lastRowKey” to the client so that the client can use this in the next rest call. I have found this to be also useful for debugging the rest output on the browser. - For Tez in particular, we need the ability to return sorted results in the rest response. In this case, results sorted based on “creation_time”. The currently existing row key in the entity table does not all for sorted order of creation time retrieval very easily. So here is proposal which incorporates some aspects of both of your proposals Varun. I think we should expose a way for frameworks like Tez to store data sorted as per their criteria. And also allow them to specify when they want to query this specially sorted data. Today, Tez wants it sorted in entity creation time. Tomorrow, that could change. Also, today some other framework like Spark might want entities sorted based on something else. So putting it in the entity table's row key becomes a tough decision. I propose we allow for auxiliary tables to be created for entities via cluster configuration settings. The auxiliary table name etc will be set in config in just like the timeline entity table name is set. This auxiliary table is specifically for entities, so has the same structure. Now, when tez’s timeline client creates a timeline entity, it will create it as it does right now but in addition, it will populate two new members of TimelineEntity object: - auxiliaryTableName which contains the desired table name - auxillaryEncodedKey which contains a byte array value of {noformat} “Inv(creation_time)!entity_id” {noformat} This is to be used as part of the row key suffix in the auxiliary table. Timeline service does not know what this byte value is, it does not care. It only adds this after the regular row key prefix of {code} “user!cluster!flow!Inv(flow run id) ! application!entitytype!” {code} Now it sends this write to timeline service. At the hbase writer side, we notice that the auxiliary table and auxiliary key are populated in the timeline entity object, so we do two writes. One write goes to our regular entity table with existing row key structure and other write goes to the auxiliary table with the row key of {code} “user!cluster!flow!Inv(flow run id)! application!entitytype!”{code}. On the reader side, we allow the rest api to now specify explicitly if the client want reads from the auxillary table. Else reads go to the regular entity table. For frameworks like Tez, whenever they need sorted data based on creation time, perhaps in their UI, they know that, so they can now specify as part of the query param in their rest query that this is for the auxiliary table. This way, we provide frameworks a way to store data in whichever sorted order they want and for them to determine queries need that sorted data. was (Author: vrushalic): Catching up on this thread. I have tried to read through all the comments and discussions on this jira but please correct me if I am mistaken. Two objectives here: 1) We are looking for a way to paginate results/response. 2) Ability to return sorted results in the rest response (sorted on something other than row key) Thoughts on these: - We are looking for a way to paginate results/response. This pagination requirement is independent of any particular framework like Tez. With hRaven, our experience has been that more often than not, we end up enabling pagination support for most APIs. So, in general, our rest api calls should support pagination. Pagination via the REST query: This involves, in a generic fashion, being able to send in a “startFromRowKey” in the rest query. Say we extend our
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15511129#comment-15511129 ] Vrushali C edited comment on YARN-5585 at 9/21/16 8:51 PM: --- Catching up on this thread. I have tried to read through all the comments and discussions on this jira but please correct me if I am mistaken. Two objectives here: 1) We are looking for a way to paginate results/response. 2) Ability to return sorted results in the rest response (sorted on something other than row key) Thoughts on these: - We are looking for a way to paginate results/response. This pagination requirement is independent of any particular framework like Tez. With hRaven, our experience has been that more often than not, we end up enabling pagination support for most APIs. So, in general, our rest api calls should support pagination. Pagination via the REST query: This involves, in a generic fashion, being able to send in a “startFromRowKey” in the rest query. Say we extend our rest apis to accept such a parameter, it becomes generic enough to fetch N rows after this particular “startFromRowKey” value. The first rest api call will not send in anything, but each rest response will return “lastRowKey” to the client so that the client can use this in the next rest call. I have found this to be also useful for debugging the rest output on the browser. - For Tez in particular, we need the ability to return sorted results in the rest response. In this case, results sorted based on “creation_time”. The currently existing row key in the entity table does not all for sorted order of creation time retrieval very easily. So here is proposal which incorporates some aspects of both of your proposals Varun. I think we should expose a way for frameworks like Tez to store data sorted as per their criteria. And also allow them to specify when they want to query this specially sorted data. Today, Tez wants it sorted in entity creation time. Tomorrow, that could change. Also, today some other framework like Spark might want entities sorted based on something else. So putting it in the entity table's row key becomes a tough decision. I propose we allow for auxiliary tables to be created for entities via cluster configuration settings. The auxiliary table name etc will be set in config in just like the timeline entity table name is set. This auxiliary table is specifically for entities, so has the same structure. Now, when tez’s timeline client creates a timeline entity, it will create it as it does right now but in addition, it will populate two new members of TimelineEntity object: - auxiliaryTableName which contains the desired table name - auxillaryEncodedKey which contains a byte array value of {code} “Inv(creation_time)!entity_id” {code}. This is to be used as part of the row key suffix in the auxiliary table. Timeline service does not know what this byte value is, it does not care. It only adds this after the regular row key prefix of {code} “user!cluster!flow!Inv(flow run id) ! application!entitytype!” {code} Now it sends this write to timeline service. At the hbase writer side, we notice that the auxiliary table and auxiliary key are populated in the timeline entity object, so we do two writes. One write goes to our regular entity table with existing row key structure and other write goes to the auxiliary table with the row key of {code} “user!cluster!flow!Inv(flow run id)! application!entitytype!”{code}. On the reader side, we allow the rest api to now specify explicitly if the client want reads from the auxillary table. Else reads go to the regular entity table. For frameworks like Tez, whenever they need sorted data based on creation time, perhaps in their UI, they know that, so they can now specify as part of the query param in their rest query that this is for the auxiliary table. This way, we provide frameworks a way to store data in whichever sorted order they want and for them to determine queries need that sorted data. was (Author: vrushalic): Catching up on this thread. I have tried to read through all the comments and discussions on this jira but please correct me if I am mistaken. Two objectives here: 1) We are looking for a way to paginate results/response. 2) Ability to return sorted results in the rest response (sorted on something other than row key) Thoughts on these: - We are looking for a way to paginate results/response. This pagination requirement is independent of any particular framework like Tez. With hRaven, our experience has been that more often than not, we end up enabling pagination support for most APIs. So, in general, our rest api calls should support pagination. Pagination via the REST query: This involves, in a generic fashion, being able to send in a “startFromRowKey” in the rest query. Say we extend our rest
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15494251#comment-15494251 ] Varun Saxena edited comment on YARN-5585 at 9/15/16 7:08 PM: - Just to summarise the suggestions given for folks to refer to. * Applications (like Tez) would know best how to interpret their entity IDs' and how they can be descendingly sorted. Most entity IDs' seem to have some sort of monotonically increasing sequence like app ID. We can hence open up a PUBLIC interface which ATSv2 users like Tez can implement to decide how to encode and decode a particular entity type so that it is stored in descending sorted fashion (based on creation time) in ATSv2. Encoding and decoding similar to AppIDConverter written in our code.Because if row keys themselves can be sorted, this will be performance wise the best possible solution. Refer to [comment | https://issues.apache.org/jira/browse/YARN-5585?focusedCommentId=15470803=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15470803] ** _Pros of the approach:_ **# Lookup will be fast. ** _Cons of the approach:_ **# We are depending on application to provide some code for this to work. Corresponding JAR will have to be placed in classpath. Folks in other projects may not be pleased to not have inbuilt support for this in ATS. **# Entity IDs' may not always have a monotonically increasing sequence like App IDs'. * We can keep another table, say EntityCreationTable or EntityIndexTable with row key as {{cluster!user!flow!flowrun!app!entitytype!reverse entity creation time!entityid}}. We will make an entry into this table whenever created time is reported for the entity. The real data would still reside in the main entity table. Entities in this table will be sorted descendingly. On read side, we can first peek into this table to get relevant records in descending fashion (based on limit and/or fromId) and then use this info to query entity table. We can do this in two ways. We can get created times from querying this index table and apply a filter of created time range. Or alternatively we can try out MultiRowRangeFilter. That from javadoc of HBase seems to be efficient. We will have to do some processing to determine these multiple row key ranges. Refer to [comment | https://issues.apache.org/jira/browse/YARN-5585?focusedCommentId=15472669=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15472669] ** _Note:_ Client should not send different created times for the same entity otherwise that will lead to an additional row. If different created time would be reported more than once we will have to consider the latest one. ** _Pros of the approach:_ **# Solution provided within ATS. **# Extra write only when created time is reported. ** _Cons of the approach:_ **# Extra peek into the index table on the read side. Single entity read can still be served directly from entity table though. * Another option would be to change the row key of entity table to {{cluster!user!flow!flowrun!app!entitytype!reverse entity creation time!entityid}} and have another table to map {{cluster!user!flow!flowrun!app!entitytype!entityid}} to entity created time. So for a single entity call (HBase Get) we will have to first peek into the new table and then get records from entity table. ** _Cons of the approach:_ **# On write side, we will have to first lookup into the index table which has the entity created time or on every write client should supply entity created time. First would impact write performance and latter may not be feasible for client to send. **# What should be the row key if client does not supply created time on first write but supplies the created time on a subsequent write. cc [~sjlee0], [~vrushalic], [~rohithsharma], [~gtCarrera9] was (Author: varun_saxena): Just to summarise the suggestions given for folks to refer to. * Applications (like Tez) would know best how to interpret their entity IDs' and how they can be descendingly sorted. Most entity IDs' seem to have some sort of monotonically increasing sequence like app ID. We can hence open up a PUBLIC interface which ATSv2 users like Tez can implement to decide how to encode and decode a particular entity type so that it is stored in descending sorted fashion (based on creation time) in ATSv2. Encoding and decoding similar to AppIDConverter written in our code.Because if row keys themselves can be sorted, this will be performance wise the best possible solution. Refer to [comment | https://issues.apache.org/jira/browse/YARN-5585?focusedCommentId=15470803=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15470803] ** _Pros of the approach:_ **# Lookup will be fast. ** _Cons of the approach:_ **# We are depending on application to provide some code for this to work. Corresponding JAR
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15472837#comment-15472837 ] Varun Saxena edited comment on YARN-5585 at 9/8/16 5:29 AM: IIUC, Tez DAG ID is a combination of YARN App ID and DAG sequence ID. Isnt this DAG sequence ID monotonically increasing and assigned to DAGs' as they are run and assigned to them in sequence ? I was assuming they were. That is why I suggested storing DAG ID as 16 bytes (8 bytes of inverted cluster timsetamp from app id + 4 bytes of inverted seq id from app id + 4 bytes of inverted DAG seq number). Padding in this case wont be required. Anyways other solutions have been proposed and we can come back to this only if necessary. Or maybe we can have both above solution and below one as well. was (Author: varun_saxena): IIUC, Tez DAG ID is a combination of YARN App ID and DAG sequence ID. Isnt this DAG sequence ID monotonically increasing and assigned to DAGs' as they are run in sequence ? I was assuming they were. That is why I suggested storing DAG ID as 16 bytes (8 bytes of inverted cluster timsetamp from app id + 4 bytes of inverted seq id from app id + 4 bytes of inverted DAG seq number). Padding in this case wont be required. Anyways other solutions have been proposed and we can come back to this only if necessary. Or maybe we can have both above solution and below one as well. > [Atsv2] Add a new filter fromId in REST endpoints > - > > Key: YARN-5585 > URL: https://issues.apache.org/jira/browse/YARN-5585 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelinereader >Reporter: Rohith Sharma K S >Assignee: Rohith Sharma K S > Attachments: YARN-5585.v0.patch > > > TimelineReader REST API's provides lot of filters to retrieve the > applications. Along with those, it would be good to add new filter i.e fromId > so that entities can be retrieved after the fromId. > Example : If applications are stored database, app-1 app-2 ... app-10. > *getApps?limit=5* gives app-1 to app-10. But to retrieve next 5 apps, it is > difficult. > So proposal is to have fromId in the filter like > *getApps?limit=5&=app-5* which gives list of apps from app-6 to > app-10. > This is very useful for pagination in web UI. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15470538#comment-15470538 ] Rohith Sharma K S edited comment on YARN-5585 at 9/7/16 12:56 PM: -- bq. Can you tell me the use case ? Listing all DAGs' or listing DAGs' within an app ? Or something else ? Typically how many DAGs' can there be per app ? Basic use case is to achieve pagination nevertheless apps/flows/entities such as DAG or containers or any other user entities or system entities. Currently limit is 100 for any entities to retrieve. Say if number of entities is 200. Then REST call retrieves 100 entities. And how to retrieve 100 to 200 entities? was (Author: rohithsharma): bq. Can you tell me the use case ? Listing all DAGs' or listing DAGs' within an app ? Or something else ? Typically how many DAGs' can there be per app ? Basic use case is to achieve pagination nevertheless DAG or containers or any other user entities. Currently limit is 100 for any entities to retrieve. Say if number of entities is 200. Then REST call retrieves 100 entities. And how to retrieve 100 to 200 entities? > [Atsv2] Add a new filter fromId in REST endpoints > - > > Key: YARN-5585 > URL: https://issues.apache.org/jira/browse/YARN-5585 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelinereader >Reporter: Rohith Sharma K S >Assignee: Rohith Sharma K S > Attachments: YARN-5585.v0.patch > > > TimelineReader REST API's provides lot of filters to retrieve the > applications. Along with those, it would be good to add new filter i.e fromId > so that entities can be retrieved after the fromId. > Example : If applications are stored database, app-1 app-2 ... app-10. > *getApps?limit=5* gives app-1 to app-10. But to retrieve next 5 apps, it is > difficult. > So proposal is to have fromId in the filter like > *getApps?limit=5&=app-5* which gives list of apps from app-6 to > app-10. > This is very useful for pagination in web UI. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15455257#comment-15455257 ] Varun Saxena edited comment on YARN-5585 at 9/1/16 12:40 PM: - So I though a little bit over it and I think there is a solution possible for fetching apps within a cluster without much of performance impact. Because this seems to be your use case. What we can do is that we can get the required App IDs' from App to flow table first as app ids' in this table are sorted and extract applicable flows from there. And then get data from the application table using these unique flows to get more specific information about the apps. We have something called MultiRowRangeFilter in HBase which can help us specify multiple row key ranges. We can only return those apps which we found from app to flow table. And from a performance viewpoint we can assume there will always be a reasonable limit specified. _Example:_ Assume, in a cluster we have applications from application_111_0001 to application_111_0034 (running or completed). These apps will be stored in a descending order in app to flow table. Let us say you want to get latest 10 apps (i.e. limit in your query is 10). What we can do is get first 10 apps from app to flow table i.e. application_111_0034 to application_111_0025. We can use PageFilter to return only first 10 records. This is the result set we can return back. Assume application IDs' ending with _0034, _0031 and _0027 belong to flow1 and rest to flow2. We can then use this info to query app table. So to get detailed info for these 10 apps in a single shot from application table, what we can do is as under : * Create a MultiRowRangeFilter * For flow1. add start row as {{cluster!user!flow1!application_111_0034}} and stop row as {{cluster!user!flow1!application_111_0027}}. We can make stop row inclusive. We can then add this start/stop row pair into the multi row range filter created. * And for flow2, start row can be {{cluster!user!flow2!application_111_0033}} and stop row as {{cluster!user!flow2!application_111_0024}}. We can then add this start/stop row pair into the multi row range filter created. This would be slower than getting all apps when flow or flow run is specified but would be faster than doing full table scan of application table, especially when it grows large. Maybe I can raise a separate JIRA for this and handle it there if this is a real use case. was (Author: varun_saxena): So I though a little bit over it and I think there is a solution possible for fetching apps within a cluster without much of performance impact. Because this seems to be your use case. What we can do is that we can get the required App IDs' from App to flow table first as app ids' in this table are sorted and extract applicable flows from there. And then get data from the application table using these unique flows to get more specific information about the apps. Say pass a flow to appids' map. We have something called MultiRowRangeFilter in HBase which can help us specify multiple row key ranges. We can only return those apps which we found from app to flow table. And from a performance viewpoint we can assume there will always be a reasonable limit specified. _Example:_ Assume, in a cluster we have applications from application_111_0001 to application_111_0034 (running or completed). These apps will be stored in a descending order in app to flow table. Let us say you want to get latest 10 apps (i.e. limit in your query is 10). What we can do is get first 10 apps from app to flow table i.e. application_111_0034 to application_111_0025. We can use PageFilter to return only first 10 records. This is the result set we can return back. Assume application IDs' ending with _0034, _0031 and _0027 belong to flow1 and rest to flow2. We can then use this info to query app table. So to get detailed info for these 10 apps in a single shot from application table, what we can do is as under : * Create a MultiRowRangeFilter * For flow1. add start row as {{cluster!user!flow1!application_111_0034}} and stop row as {{cluster!user!flow1!application_111_0027}}. We can make stop row inclusive. We can then add this start/stop row pair into the multi row range filter created. * And for flow2, start row can be {{cluster!user!flow2!application_111_0033}} and stop row as {{cluster!user!flow2!application_111_0024}}. We can then add this start/stop row pair into the multi row range filter created. This would be slower than getting all apps when flow or flow run is specified but would be faster than doing full table scan of application table, especially when it grows large. Maybe I can raise a separate JIRA for this and handle it there if this is a real use case. > [Atsv2] Add a new filter fromId in
[jira] [Comment Edited] (YARN-5585) [Atsv2] Add a new filter fromId in REST endpoints
[ https://issues.apache.org/jira/browse/YARN-5585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=15452583#comment-15452583 ] Varun Saxena edited comment on YARN-5585 at 8/31/16 3:51 PM: - If the use case is only for apps then the row keys in application table are stored in sorted manner (in descending order) within the scope of a flow / flow run. And we can easily support fromId alongwith limit to achieve some sort of pagination here without any performance penalty. However, the problem with this kind of an approach is that new apps keep on getting added so result may not be latest. For instance, if there are 100 apps app100-app1 in ATS and we show 10 apps on each page. Then, if we move to page 3 we will show apps from app80-app71 but it is possible that say 5 more apps get added in the meantime i.e. we now have app105 to app1 in ATS. Ideally page 3 should then show app85-app76. But I guess this would have already been considered. Entities in entity table though are not sorted because entity could be anything. If we have a similar use case for containers, we can consider separating it out to a different table and have special handling for it. But there should be a use case for it. was (Author: varun_saxena): If the use case is only for apps then the row keys in application table are stored in sorted manner (in descending order) within the scope of a flow / flow run. And we can easily support fromId alongwith limit to achieve some sort of pagination here without any performance penalty. However, the problem with this kind of an approach is that new apps keep on getting added so result may not be latest. For instance, if there are 100 apps app100-app1 in ATS and we show 10 apps on each page. Then, if we move to page 3 we will show apps from app80-app71 but it is possible that say 5 more apps get added in the meantime i.e. we not have app105 to app1 in ATS. Ideally page 3 should then show app85-app76. Entities in entity table though are not sorted because entity could be anything. If we have a similar use case for containers, we can consider separating it out to a different table and have special handling for it. But there should be a use case for it. > [Atsv2] Add a new filter fromId in REST endpoints > - > > Key: YARN-5585 > URL: https://issues.apache.org/jira/browse/YARN-5585 > Project: Hadoop YARN > Issue Type: Sub-task > Components: timelinereader >Reporter: Rohith Sharma K S >Assignee: Rohith Sharma K S > > TimelineReader REST API's provides lot of filters to retrieve the > applications. Along with those, it would be good to add new filter i.e fromId > so that entities can be retrieved after the fromId. > Example : If applications are stored database, app-1 app-2 ... app-10. > *getApps?limit=5* gives app-1 to app-10. But to retrieve next 5 apps, it is > difficult. > So proposal is to have fromId in the filter like > *getApps?limit=5&=app-5* which gives list of apps from app-6 to > app-10. > This is very useful for pagination in web UI. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org