RAIB

(c)'2007 by Andreas Filsinger

Redundant Array of Independent dataBases

The goal of RAIB (for firebird 3.0) is to ensure the availability of a firebird database - even if hardware fails. This is done by fail-over servers (spares) or Log-Packets (.flog File) on a remote Filesystem. In other Environments (without RAIB) you loose all Data from the Moment of your last Backup until the Moment of faliure. With RAIB you are able to get your Database up to the state where it was. This is just a concept - BUT implementing this ideas will make firebird 3.0 (with the vulcan engine) unbreakable.
Remember that malfunction of a single transistor in your high-end-server will bring firebird down (And modern servers have millions of it). Maybe your HD is OK, or you trust on a RAID 5 Storage. But the Question is: How long does it take to get the firebird-service online again!
With RAIB you can immediatly connect your application to a spare-Host after the failure. In this Moment you are restricted to "read-onyl" if you reduce to "read only" SQL-Statements, a admin can switch the spare to the master mode, so all the functionality of the firebird database can be used.

How it will work

The work is done by 6 (new?) techniques, not yet implemented in firebird:

1. Heart-Beat-Sequence: a BIGINT living inside the database, counts the impacts to the database since the Statement "CREATE DATABASE". This will help naming Statement-Log-Packets
   
2. Statement-Classification: by evaluating and executing a statement it is classified as a "R"- "W"- or "I"-Statement.
   
3. Statement-Log-Packets: Logs DDL/DML "W"-Statements in a way so they can build basis of a Statement - "playback".
   
4. snap-shot-backups: a Backup-Mode where all the impact AFTER the Backups was started are invisible to the Backup.
   
5. Statement-Playback: a system where server can execute "Statement-Packets".
6. Playback-Mode: gbak sets a database in this mode when doing statement Playback to a database.

Heart-Beat-Sequence

SEQ_HEART_BEAT

A Database has an internal GENERATOR (the Heart-Beat-Sequence = HBS [BIGINT]) that is incremented after each statement commit doing some change to data or metadata. A read-Only Statement should not incrememt the HBT.

Statement Classification

firebird executes SQL-Code like "DROP TABLE SAMPLE" or Service Requests like "BACKUP DATABASE". If we want something like a spare-server we should filter the requests that really do some change to the Database. The three classes are "W"rite / "R"ead and "I"nformal.

"W" Requests

"R" Requests

"I" Requests

Samples
Client-Request: DROP TABLE CLUB$2873
Server-Classification: "W" / SEQ_HEART_BEAT=928372

Client-Request: SELECT CURRENT_TIMESTAMP from RDB$DATABASE
Server-Classification: "I"

Client-Request: SELECT COUNT(RID) from HUGETABLE
Server-Classification: R

Statement-Log-Packets

A SQL-Statement like "DROP TABLE HUGEONE" may purge thousands of Pages of the Database File, and modify 1 or 2. The Log-Packet itself schould NOT contain all the changed Database Pages or such low-level Details but the SQL-Statement itself. We can trust on the fact that the spares have all the knowledge to understand the Statement and execute it exactly like the master. This information is transmitted to one or more spares. This information is also stored in the Database itself for a time (not part of a backup). At least until its prooven that a n.fabk is available. After the Log-DataBlock must succsessfully

1. Do the Statement, detect while doing it, if it is harmfull to the Database (W oder R)
2. if W: compile a log-block with the name GEN_ID(HBT,1); if R: break
3. Send the LOG over the line.

Generating a LOG-Packet

"R"-Statements /"W"-Statements

Each Request, on its way through the Database holds a Flag, saying "I am Read/Only" or "I did some impact" to the Database. This Flag (the R/W-Flag) is set by the engine in the actual context, because you can formulate a statement wich is 99% a "R"-one, but there comes a day, this statement do a change to the Database. When the engine detects a change to Data or Metadata it sets the "Write"-Flag, this statement was prooven now to have a impact to the database, it is a "W"-Statement and must at its End increment the SEQ_HEART_BEAT and a Statement-Log is written.

The Log-Packet is assembled by the engine step by step right in the moment by parsing and executing it. The LOG-Packet is the "compiled" version of the Statement. The LOG-Paket is written to a memory block. If a "W" Statement reads CURRENT_TIMESTAMP (Timestamp) or RND (Randomize) a copy of the returned value MUST be stored in the LOG, because the spares can not trust on own values. If a Request is fully executed, and the engine didn't set the "W"-Flag - the LOG-Packet is purged, because it did no harm to the database it can say us nothing about history. Else (a "W"-Statement) the Packet is populatetd to the spares. The server itself hold a persistent copy of the Log-Packet to help spare get up to date if they missed something. Only a successfull backup can delete the Log-Packets that are generated before the .fbak HBS.

Quasi "W" Statements

select CURRENT_TIMESTAMP from RDB$DATABASE

is a "W"-Statement that can only be executed by the master, because the client ask for information that is NOT part of the Database itself. So all external Information going throgh the engine are Quasi-"W" (user account infos, or coming user SQL Statements). They MUST be executed by the master, but do NOT goto the Log.

Statement

UPDATE PERSON SET MOMENT = CURRENT_TIMESTAMP WHERE LOAN = RND(2920)

Log-Paket

P:RND=2093 P:CURRENT_TIMESTAMP P:CURRENT_TIMESTAMP P:CURRENT_TIMESTAMP S:UPDATE PERSON ...

Statement-Playback

Database in spare mode

• the LOG-Receive Thread

the master initiates communication to a spare and post LOG Packets to the spare. Alle Packets are flushed into memory - if CRC is ok, the spare sends a ACK to the master. The master do NOT wait until the spare executes the statement because the master already prooved that the statement is executable.

• the Worker-Thread

read the LOG-Packets, and execute it against the own Database.

• LOG-missed

if the spare was offline a time, she sees that local.HBT <> LOG.HBT+1. Before it can execute actual LOG-Packets she MUST ask the master (or other spares) for the missing LOG Packets. After doing the old stuff it can start to do execute the actual receiced LOG-Packtes. While this time, the log-Receive thread isnt stopping filling the LOG-Buffer.

Playback-Mode

The Playback-Mode (or Spare Mode) of a database is a new type of Database "Shutdown" Mode. Connections to this host in Spare Mode see no difference to a normal Database. work normal but refuse "W" Statement with a special error message. They may redirekted to a master.

be a local spare

the master writes the LOG-Packets (FileName <HBT>.flog) to a (remote) filesystem.

be a cold spare

a firebird-server receives the log-Packets but do not execute them. The worker-Thread writes the .flog Files to the filesystem if the system is idle.

be a hot spare

a firebird-server receives and execute the log-Packets a fast as possible.

be a master

planned switiching the role from master to spare

1. the master do not accept W Statements any more. (Try again in 2 Minutes!)
2. the master ensures that one spare is ready to take over (master.HTB=spare.HTB)
3. the master switch to "spare" Mode
4. the spare switch to "master" Mode

Database in master mode

no limitations. He must bring the Log-Packets to the world:

• By UDP to the spares.
• By fopen to the filesystem.
• By move to the memory.

(add a ) RDB$Cluster Table

inside the database is the information about the spares and the master.

IP               Role       last-HBT
192.168.115.192  master
192.168.115.196  spare
192.168.115.80   spare
30.23.12.3       spare

detecting the failure

• Master is Not Responding
• Spare say they have problems executing crazy Packets.

Positive Side Effekts

if a client assumes that a special statement is "R" AND HBT of a spare is the HBT of the master she can post this statement to a spare to take load away from the master. The spare will bring the same result as the master.

Communication master<->spare

if a spare comes up (on connection request of a master), a spare can ask the master how

Good to have several spares

it is enough fore the master to have only one responding spare. If a connection to a spare fails the master retries it every 20 Seconds.

Restore Szenario

1. restore the database from a .fbak
2. read the value of GEN_HBT
3. let the server execute all *.flog Files beginning from GEN_HBT+1

New ERROR Messages

"W"-Statement not allowed on a spare

the client must understand a new error-Msg when trying to place a statement to a spare that was a master before. the client must interrpret the new connection string and (re)prepare a open statements.

1. Error-Msg with a Retry-Option: "Im a Spare, please retry with [%server:%database]"

Failed to Log your Statement

if the master can not log the statement to any spare this error MSG is generated.

the lying master concept

Imaging a host wich is a spare (in truth) and tells the world that she is a Master. This is a spare wich executes all the "R" Statement, but automatically routes "W"-Statements to a master without telling the connection something about it. The result is routed back to the connection AFTER she has executed the Log-Packet, so that this spare is "true" wich its single connection. There may be a "gab" of Log-Packets (more to come here).
You may have a Grid of "lying masters" in a webshop scenario. Alle the "W" Work is done by one hidden master, not slowed down by anoing "R" Statements of any connection.

the lazy worker

• say "Im done" if all checks are done that ensure "T" will succseed, but dont do the work!

the Cash Assoziator

• Cashing: is a technique that repeats a saved answer to an known question without "understanding" the question again. This is Quick!
• Trick 1: Always use cash if since last "R" there was no "T"
• Trick 2: On first insert in cash, build a dependecy-tree, on what the Answers depends on, if asked again, only check tree, if not more developed than the saved version -> use cashed answer
• Trick 3: Cash-Questions are parallel to Answers, we have an unlimited Count of CPUs.

the Context Bubble

• repesents the Dependecy Tree

A cash Element contains the "cas-Azzoziator-ID" the "Answer" the "Context-Bubble". In "idle" Time the Cash is cleared off all the Entries wich Context-Bubbles blobs. A "R" Statement can be cashed if "Context" Bubble still is alive

the new I/O Model

• extends the classic I/O Modell by the possibility to Ask the I for a List of dependencies of the I

Project names alternatives

"firebird D" (Firebirdie)

GenerationsCount: SQL Statement, das mich dazu gemacht hat.