Re: [HACKERS] GSoC 2014 proposal
On Wed, Apr 2, 2014 at 2:22 PM, Alexander Korotkov aekorot...@gmail.comwrote: On Tue, Apr 1, 2014 at 2:23 PM, Heikki Linnakangas hlinnakan...@vmware.com wrote: The BIRCH algorithm as described in the paper describes building a tree in memory. If I understood correctly, you're suggesting to use a pre-built GiST index instead. Interesting idea! There are a couple of signifcant differences between the CF tree described in the paper and GiST: 1. In GiST, a leaf item always represents one heap tuple. In the CF tree, a leaf item represents a cluster, which consists of one or more tuples. So the CF tree doesn't store an entry for every input tuple, which makes it possible to keep it in memory. 2. In the CF tree, all entries in a leaf node must satisfy a threshold requirement, with respect to a threshold value T: the diameter (or radius) has to be less than T. GiST imposes no such restrictions. An item can legally be placed anywhere in the tree; placing it badly will just lead to degraded search performance, but it's still a legal GiST tree. 3. A GiST index, like any other index in PostgreSQL, holds entries also for deleted tuples, until the index is vacuumed. So you cannot just use information from a non-leaf node and use it in the result, as the information summarized at a non-leaf level includes noise from the dead tuples. Can you elaborate how you are planning to use a GiST index to implement BIRCH? You might also want to take a look at SP-GiST; SP-GiST is more strict in where in the tree an item can be stored, and lets the operator class to specify exactly when a node is split etc. Hmmm, it's likely I've imagined something quite outside of this paper, and even already suggested it to Ivan... :) I need a little time to rethink it. Using GiST we can implement BIRCH-like clustering like so: 1) Build a CF tree as GiST index without restriction of T threshold value. 2) Scan CF tree with threshold T with some auxiliary operator. If consistent method see CF entry which diameter is greater than T then it returns true. Otherwise it returns false and put this CF entry into output area (could be either in-memory or temporary table). 3) Process other steps of algorithm as usual. This modification would have following advantages: 1) User can build GiST index once and then try clustering with different parameters. Initial GiST index build would be slowest operation while other steps is expected to be fast. 2) Use GiST infrastructure and automatically get buffering build. The drawback is that building GiST index is more expensive than building in-memory CF tree with given threshold T (assuming T is well chosen). Does it make any sense? -- With best regards, Alexander Korotkov.
Re: [HACKERS] GSoC 2014 proposal
On 04/03/2014 04:15 PM, Alexander Korotkov wrote: On Wed, Apr 2, 2014 at 2:22 PM, Alexander Korotkov aekorot...@gmail.comwrote: On Tue, Apr 1, 2014 at 2:23 PM, Heikki Linnakangas hlinnakan...@vmware.com wrote: The BIRCH algorithm as described in the paper describes building a tree in memory. If I understood correctly, you're suggesting to use a pre-built GiST index instead. Interesting idea! There are a couple of signifcant differences between the CF tree described in the paper and GiST: 1. In GiST, a leaf item always represents one heap tuple. In the CF tree, a leaf item represents a cluster, which consists of one or more tuples. So the CF tree doesn't store an entry for every input tuple, which makes it possible to keep it in memory. 2. In the CF tree, all entries in a leaf node must satisfy a threshold requirement, with respect to a threshold value T: the diameter (or radius) has to be less than T. GiST imposes no such restrictions. An item can legally be placed anywhere in the tree; placing it badly will just lead to degraded search performance, but it's still a legal GiST tree. 3. A GiST index, like any other index in PostgreSQL, holds entries also for deleted tuples, until the index is vacuumed. So you cannot just use information from a non-leaf node and use it in the result, as the information summarized at a non-leaf level includes noise from the dead tuples. Can you elaborate how you are planning to use a GiST index to implement BIRCH? You might also want to take a look at SP-GiST; SP-GiST is more strict in where in the tree an item can be stored, and lets the operator class to specify exactly when a node is split etc. Hmmm, it's likely I've imagined something quite outside of this paper, and even already suggested it to Ivan... :) I need a little time to rethink it. Using GiST we can implement BIRCH-like clustering like so: 1) Build a CF tree as GiST index without restriction of T threshold value. 2) Scan CF tree with threshold T with some auxiliary operator. If consistent method see CF entry which diameter is greater than T then it returns true. Otherwise it returns false and put this CF entry into output area (could be either in-memory or temporary table). 3) Process other steps of algorithm as usual. I still don't understand how that would work. You can't trust the non-leaf entries, because their CF value can contain deleted entries. So you have to scan every tuple anyway. Once you do that, what's the point of the index? - Heikki -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] GSoC 2014 proposal
On Thu, Apr 3, 2014 at 11:21 PM, Heikki Linnakangas hlinnakan...@vmware.com wrote: On 04/03/2014 04:15 PM, Alexander Korotkov wrote: On Wed, Apr 2, 2014 at 2:22 PM, Alexander Korotkov aekorot...@gmail.com wrote: On Tue, Apr 1, 2014 at 2:23 PM, Heikki Linnakangas hlinnakan...@vmware.com wrote: The BIRCH algorithm as described in the paper describes building a tree in memory. If I understood correctly, you're suggesting to use a pre-built GiST index instead. Interesting idea! There are a couple of signifcant differences between the CF tree described in the paper and GiST: 1. In GiST, a leaf item always represents one heap tuple. In the CF tree, a leaf item represents a cluster, which consists of one or more tuples. So the CF tree doesn't store an entry for every input tuple, which makes it possible to keep it in memory. 2. In the CF tree, all entries in a leaf node must satisfy a threshold requirement, with respect to a threshold value T: the diameter (or radius) has to be less than T. GiST imposes no such restrictions. An item can legally be placed anywhere in the tree; placing it badly will just lead to degraded search performance, but it's still a legal GiST tree. 3. A GiST index, like any other index in PostgreSQL, holds entries also for deleted tuples, until the index is vacuumed. So you cannot just use information from a non-leaf node and use it in the result, as the information summarized at a non-leaf level includes noise from the dead tuples. Can you elaborate how you are planning to use a GiST index to implement BIRCH? You might also want to take a look at SP-GiST; SP-GiST is more strict in where in the tree an item can be stored, and lets the operator class to specify exactly when a node is split etc. Hmmm, it's likely I've imagined something quite outside of this paper, and even already suggested it to Ivan... :) I need a little time to rethink it. Using GiST we can implement BIRCH-like clustering like so: 1) Build a CF tree as GiST index without restriction of T threshold value. 2) Scan CF tree with threshold T with some auxiliary operator. If consistent method see CF entry which diameter is greater than T then it returns true. Otherwise it returns false and put this CF entry into output area (could be either in-memory or temporary table). 3) Process other steps of algorithm as usual. I still don't understand how that would work. You can't trust the non-leaf entries, because their CF value can contain deleted entries. So you have to scan every tuple anyway. Once you do that, what's the point of the index? Yeah, it is limitation of this idea. It's not going to be auto-updatable CF. User can build index on freshly vacuumed table and play with clustering some time. Updates on table during that time would be either allowed clustering error or prohibited. Another potential solution is to let this index to hold some snapshot of data. But it seems not possible to do in extension now. -- With best regards, Alexander Korotkov.
Re: [HACKERS] GSoC 2014 proposal
On Tue, Apr 1, 2014 at 2:23 PM, Heikki Linnakangas hlinnakan...@vmware.comwrote: The BIRCH algorithm as described in the paper describes building a tree in memory. If I understood correctly, you're suggesting to use a pre-built GiST index instead. Interesting idea! There are a couple of signifcant differences between the CF tree described in the paper and GiST: 1. In GiST, a leaf item always represents one heap tuple. In the CF tree, a leaf item represents a cluster, which consists of one or more tuples. So the CF tree doesn't store an entry for every input tuple, which makes it possible to keep it in memory. 2. In the CF tree, all entries in a leaf node must satisfy a threshold requirement, with respect to a threshold value T: the diameter (or radius) has to be less than T. GiST imposes no such restrictions. An item can legally be placed anywhere in the tree; placing it badly will just lead to degraded search performance, but it's still a legal GiST tree. 3. A GiST index, like any other index in PostgreSQL, holds entries also for deleted tuples, until the index is vacuumed. So you cannot just use information from a non-leaf node and use it in the result, as the information summarized at a non-leaf level includes noise from the dead tuples. Can you elaborate how you are planning to use a GiST index to implement BIRCH? You might also want to take a look at SP-GiST; SP-GiST is more strict in where in the tree an item can be stored, and lets the operator class to specify exactly when a node is split etc. Hmmm, it's likely I've imagined something quite outside of this paper, and even already suggested it to Ivan... :) I need a little time to rethink it. -- With best regards, Alexander Korotkov.
Re: [HACKERS] GSoC 2014 proposal
On 03/30/2014 11:50 PM, Иван Парфилов wrote: The implementation of this algorithm would be for data type cube and based on GiST. The key concept of BIRCH algorithm is clustering feature. Given a set of N d-dimensional data points, the clustering feature CF of the set is defined as the triple CF = (N,LS,SS), where LS is the linear sum and SS is the square sum of data points. Clustering features are organized in a CF tree, which is a height balanced tree with two parameters: branching factor B and threshold T. Because the structure of CF tree is similar to B+-tree we can use GiST for implementation [2]. The GiST is a balanced tree structure like a B-tree, containing key, pointer pairs. GiST key is a member of a user-defined class, and represents some property that is true of all data items reachable from the pointer associated with the key. The GiST provides a possibility to create custom data types with indexed access methods and extensible set of queries. The BIRCH algorithm as described in the paper describes building a tree in memory. If I understood correctly, you're suggesting to use a pre-built GiST index instead. Interesting idea! There are a couple of signifcant differences between the CF tree described in the paper and GiST: 1. In GiST, a leaf item always represents one heap tuple. In the CF tree, a leaf item represents a cluster, which consists of one or more tuples. So the CF tree doesn't store an entry for every input tuple, which makes it possible to keep it in memory. 2. In the CF tree, all entries in a leaf node must satisfy a threshold requirement, with respect to a threshold value T: the diameter (or radius) has to be less than T. GiST imposes no such restrictions. An item can legally be placed anywhere in the tree; placing it badly will just lead to degraded search performance, but it's still a legal GiST tree. 3. A GiST index, like any other index in PostgreSQL, holds entries also for deleted tuples, until the index is vacuumed. So you cannot just use information from a non-leaf node and use it in the result, as the information summarized at a non-leaf level includes noise from the dead tuples. Can you elaborate how you are planning to use a GiST index to implement BIRCH? You might also want to take a look at SP-GiST; SP-GiST is more strict in where in the tree an item can be stored, and lets the operator class to specify exactly when a node is split etc. We need to implement it to create GiST-based CF-tree to use it in BIRCH algorithm. *Example of usage(approximate):* create table cube_test (v cube); + insert into cube_test values (cube(array[1.2, 0.4]), cube(array[0.5, -0.2]), cube(array[0.6, 1.0]),cube(array[1.0, 0.6]) ); create index gist_cf on cube_test using gist(v); --Prototype(approximate) --birch(maxNodeEntries, distThreshold, distFunction) SELECT birch(4.1, 0.2, 1) FROM cube_test; cluster | val1 | val2 -+--+ 1 | 1.2 | 0.4 0 | 0.5 | -0.2 1 | 0.6 | 1.0 1 | 1.0 | 0.6 Accordingly, in this GSoC project BIRCH algorithm for data type cube would be implemented. From the example, it seems that birch(...) would be an aggregate function. Aggregates in PostgreSQL currently work by scanning all the input data. That would certainly be a pretty straightforward way to implement BIRCH too. Every input tuple would be passed to the the so-called transition function (which you would write), which would construct a CF tree on-the-fly. At the end, the result would be constructed from the CF tree. With this approach, the CF tree would be kept in memory, and thrown away after the query. That would be straightforward, but wouldn't involve GiST at all. To use an index to implement an aggregate would require planner/executor changes. That would be interesting, but offhand I have no idea what that would look like. We'll need more details on that. - Heikki -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Re: [HACKERS] GSoC 2014 proposal
On 03/30/2014 11:50 PM, Иван Парфилов wrote: * Quantifiable results* Adding support of BIRCH algorithm for data type cube Aside from the details of *how* that would work, the other question is: Do we want this in contrib/cube? There are currently no clustering functions, or any other statistical functions or similar, in contrib/cube. Just basic contains/contained/overlaps operators. And B-tree comparison operators which are pretty useless for cube. Do we want to start adding such features to cube, in contrib? Or should that live outside the PostgreSQL source tree, in an separate extension, so that it could live on its own release schedule, etc. If BIRCH goes into contrib/cube, that's an invitation to add all kinds of functions to it. We received another GSoC application to add another clustering algorithm to the MADlib project. MADlib is an extension to PostgreSQL with a lot of different statistical tools, so MADlib would be a natural home for BIRCH too. But if it requires backend changes (ie. changes to GiST), then that needs to be discussed on pgsql-hackers, and it probably would be better to do a reference implementation in contrib/cube. MADlib could later copy it from there. - Heikki -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
