On 13 January 2012 12:56, Nils Breunese <[email protected]> wrote: > Matt <[email protected]> wrote: > >> Suppose we buy an sell an awful lot of fruit, mostly apples and oranges but >> sometimes bananas, pears and anything else that shoppers want. We take a >> feed (from another system) of every trade and load them all into couch as we >> go along, then define a view to aggregate them up by various different >> aspects of the sale. We can control the grouping level to produce many >> different reports out the same view and it works very well for us. Some of >> the attributes on the trades documents are not on the original feed so we >> enrich the data before PUTting them in couch. This is also fine. >> >> Unfortunately, sometimes the enrichment changes. For example each trade has >> a fruit field and we have a mapping from fruit to fruit type. Today apples, >> oranges and pears are classified as round fruit whereas bananas are long >> fruit so we can put a fruit type field on the document as we load it in but >> tomorrow that classification might change and pears might become lumpy fruit >> instead of round fruit. We could go back and update every pear trade in the >> database but we'd rather not have fruit type on the document at all and >> maintain a separate mapping "table". >> >> We thought that perhaps the map function could reference another database >> that was keyed on fruit and had documents containing the fruit type but >> don't seem to be able to perform this lookup in the mapping. We could get >> the document itself on the result but not extract just one field to augment >> the original documents. We also considered generating the view's design >> dynamically out of a mapping table and replicating the mapping into explicit >> logic in the javascript but this seemed like an abuse. >> >> In short, how would you implement a pre-aggregation lookup table? > > I think I'd just put the lookup table in the map function. Another option is > indeed to update the documents and change the fruit type. Which of these two > yields best performance probably depends on how large your lookup table would > be and how many documents you'd have to update. Map functions need to be > completely standalone in order to ensure things keep working when you start > doing things like distributing your data, so calls to external stuff is a > no-go. > > Nils. > ------------------------------------------------------------------------ > VPRO www.vpro.nl > ------------------------------------------------------------------------
Its worth noting that changing the map functions in either way will require the view to be rebuilt before it can be used which might be an expensive or timeconsuming operation. This is a common workaround for reducing the operational impact of doing that http://wiki.apache.org/couchdb/How_to_deploy_view_changes_in_a_live_environment I'd not update the docs directly as this will increase the DB size, requiring more frequent compaction. In comparison a view generation would likely be less expensive. A+ Dave
