Technical point of view: "Thinking Exadata" My opinion

This article is for IBM system architect Mr. Wang Wenjie (valen_won@hotmail.com) in its blog park Zhiding Bowen thinking EXADATA (link original address: http://www.cnblogs.com/wenjiewang/archive/2012/10/07/ 2714406.html) mentioned some of the views on the Exadata from the technical point of view gives my personal opinion of some different, of course, I have limited level, it is inevitable that omissions or mistakes.

Here's a brief summary of what Mr. Wenjie mentioned in Bowen about Exadata:

1. Data Warehouse types of applications fail to take full advantage of the features of smart scan, and Exadata can not optimize, especially for the star transformations common in data warehousing;

2. The Exadata bugs are numerous and some of the new features come in handy and illustrate some of the bugs they encountered with Bloom filters.

3. The Oracle Database has many bugs and uses Exadata. There is a big risk to the financial industry that has complicated business logic and data correctness.

4. To maintain high costs, to use Exadata, DBA need to re-learn a large number of host, storage, network knowledge, or can not be competent to work as a machine administrator;

5. For data warehouses with hundreds of fields in a single table, Exadata's Storage Index is a shame because only eight columns are automatically maintained for each table,

6. Exadata or RAC, RAC share everything architecture leads to a large number of cache fusion contention, and OLTP applications are out of tune.

RAC poor support for ERP, resulting in many ERP users do not use RAC, Exadata only provide RAC mode.

Exadata disk capacity is too large, this is simply a waste for OLTP;

9. Exadata does not provide any virtualization technology and can not take full advantage of its hardware resources, while its competitors do provide very sophisticated virtualization solutions;

10. (This year, to get together out of ten are embarrassed to say hello to people) Exadata's price is very expensive, ordinary users simply can not afford.

The following is my personal response, does not represent the official position of Oracle. Oracle's salary has not yet come to my impulse to do Oracle 5 hair. If you love to see soft paper, then you can please move Loveunix and AIXchina, because there are more. A lot of technical details is difficult to explain a few words clearly, limited to space, where to seek an introduction or a passing.

1. In fact, smart scan can be described as the core of all Exadata technologies, leaving the smart scan, Exadata has no soul. The Exadata smart scan conditions are too harsh, has been much criticized by competitors, this is also a fact. However, the article mentions that "if our report is not go FULL TABLE SCAN, you can not take advantage of this feature. Complex queries, such as Joins, sorts, group-bys, aggregation are likely to not be able to take advantage of intelligent scanning." A statement is not accurate. I am tired of enumerating here the current conditions of the smart scan :( Although this is a false "truth")

Full scans - Table, Partition, Materialized View, Index (FAST FULL SCAN Only)

Direct Path Reads

Exadata Storage

For more complex sorting, aggregation class operation, storage index has its useless. As for the star conversion, I am afraid that the author is not the fact that this article and this article details in the data warehouse, Oracle internal how to optimize the star transform some details.

Bloom filter is a hash algorithm for handling large amounts of data. The specific algorithm can refer to Wikipedia entry Bloom Filter, or refer to Google's Mr. Wu Jun's "Beauty of Mathematics" book. Of course, the bug mentioned here is not accurate, here are two bugs: 9124206 and bug: 8361126, in fact, the same bug that the base bug 8361126.

Bloom filter bug mentioned in the article should be Bug: 12637294, but this bug has been fixed in 11.2.0.3 BP11.

Another interesting point is that smart scan also uses Bloom Filter algorithm to filter data.

3. Oracle Bug is a well-known fact, from each Patchset Release / PSU bug list you can see that many of the hazards of the bug is also very large. Even the wrong result by the author is entirely a fact, but this is not without reason; most of these bugs are triggered in extreme cases. If the application has been fully tested, then rarely encountered wrong results. Trigger wrong results bug Some of the more common cases are parallel, complex table joins and other operations. MOS has a document detailing how to diagnose and analyze such problems: Wrong Results Issues - Recommended Actions [ID 150895.1]. By the way: More and more financial industry customers regard Oracle Database as a core.

4. Maintenance costs of the problem. Maintenance is not as serious as the author said. Host is a PC server, the hardware is nothing special. Operating system is Linux X86_64, many SA / DBA are already very familiar with. Network maintenance does not require additional knowledge, just know some common infiniband / cisco switch operation. Database maintenance on Exadata is not different from a normal RAC database. The only thing that needs to be relegated is storage-side knowledge, and much of this content is available on the internet. (In case it really can not do it, Oracle launched a one-stop platinum service, users can outsource management to Oracle, laugh, please enter automatically ignore advertising mode)

5. Storage Index It is a fact that only 8 columns are automatically maintained per table, but this is not a technical limitation. The Storage Maps and Netezza's Zone Maps technologies are fundamentally different. An important concept of the Storage Index is that it only works on sorted fields, which is not useful for out-of-order fields, so Storage Index's more than 8 columns per table do not do much in performance because of a table core that is needed for Sort the field is not much.

6. This issue is actually still a share of the battle between share disk and share nothing, the old topic, and not much practical significance.

7. The current SAP ERP on Oracle DB is far more than the ERP running on DB2, and is interested in viewing gartner's statistics.

8. Now cabbage price of the hard disk, a single plate on the 2-3T, and who cares about such a little space? Moreover, the amount of OLTP application data in more than 1T is also a small number.

9. This article says the truth, however

Vmware such virtualization platform currently does not pass Oracle certification;

IBM LPAR is not strictly a virtualization technology;

· Exadata can be isolated through resources like IORM / instance cage / cgroups;

The future should consider using Oracle's own OVM.

10. Compared to the high-end host + high-end storage hundreds of millions, Exadata cost is not bad, right? Now Exadata X3 launched 1/8 with, began to grab their own brother ODA's job. . .