Unstoppable Stateful PHP Web Services

Transcription

1 Max-Planck Institute for Informatics, Saarbrücken, Germany Databases and Information Systems Unstoppable Stateful PHP Web Services German Shegalov, Gerhard Weikum, and Klaus Berberich funded by WISE'06, Wuhan, China 1

2 Outline Problem Statement and Background Interaction Contracts (IC) Framework Contract between Web Services Contract between User & Browser Implementation & Experiments: Exactly-Once Web Service (EOS) Summary WISE'06, Wuhan, China 2

3 E-Business Scenario Please review and place your order Place your order Your server command (process id #20) has been terminated. Re-run your command (severity 13) in /export/home/www/your-reliable-eshop.biz/mb_1300_db.mb1 WISE'06, Wuhan, China 3

4 Transactional Recovery Insufficient Web Client Web Application Server Database Server Timeline Purchase Request Order Confirmation Start Transaction SQL Request SQL Response SQL Request SQL Response Commit Transaction ACK Transaction Restart Purchase Request Resubmission Non-idempotent execution! Atomic = At-most-once Exactly-once Idempotence needs testability, but testable state all but simple WISE'06, Wuhan, China 4

5 Real-World n-tier App Client Complicated enough? Business transactions between peers in community apps (Skype, MSN, ). Web Server Expedia Expedia App Server Sabre Sabre App Server Amadeus Amadeus App Server DB 1 DB 2 DB 3 DB 4 WISE'06, Wuhan, China 5

6 Problem Statement Application-tailored solutions for high-end e-business not available to masses of low-end service & mashup programmers Ease programming by providing generic solution in Web-service middleware Should mask all (but inevitably disruptive) failures to application program(mer mer)s: message, process, data failures WISE'06, Wuhan, China 6

7 Outline Problem Statement and Background Interaction Contracts (IC) Framework Contract between Web Services Contract between User & Browser Implementation & Experiments: Exactly-Once Web Service (EOS) Summary WISE'06, Wuhan, China 7

8 Interaction Contracts Framework Interaction Contracts Framework [Barga et al.: TOIT 04] Components and Guarantees Persistent Pcom: Persistent, testable state and messages External Xcom (e.g., humans): guarantees Bilateral Interaction Contracts Xcom Pcom = External IC (XIC) Pcom Pcom = Committed IC (CIC) Composing IC s s for Entire System: Exactly-Once Semantics App programs don t t need to handle failures WISE'06, Wuhan, China 8

9 Persistent Component Design Redo Log & Recovery Managers Piecewise determinism + Logging = Full Determinism Deterministic replay recovers Pcom's Installation Points speed up replay Failure model Crashes Message losses Malicious manipulations Disk corruption (sufficient redundancy) [ICDE 02] Transient failures due to nondeterministic Heisenbugs WISE'06, Wuhan, China 9

10 Outline Problem Statement and Background Interaction Contracts (IC) Framework Contract between Web Services Contract between User & Browser Implementation & Experiments: Exactly-Once Web Service (EOS) Summary WISE'06, Wuhan, China 10

11 CIC Principles CIC sender (Pcom1) obligations Persist state before send Tag message with a MSN Resend on timeout until stable ack Resend on receiver's inquiry Forget interaction on installed ack CIC receiver (Pcom2) obligations Eliminates duplicates by MSN Persists interaction before stable ack Inquires msg body if not in local log Ensures autonomous recovery before installed ack Weaker than using persistent queue or installing state for each interaction Pcom1 Pcom2 WISE'06, Wuhan, China 11

12 Statechart for CIC Sender CIC_SNDR_SC MSG_LOOKUP MSG_RECOVERED_TM/ SEND_MSG GET_MSG_OK SNDR_MSG_TM and not (STABLE_OK or INSTALLED_OK)/ SEND_MSG STABLE_OK STABLE_S INSTALLED_OK/ SNDR_LAST_LOGGED:='INSTALLED' SNDR_STABLE_TM and not (INSTALLED_OK or GET_MSG_OK)/ IS_INSTALLED SENDING INSTALLED_OK/ SNDR_LAST_LOGGED:='INSTALLED' INSTALLED_S SNDR_ND/ SEND_MSG PREPARE_PERSISTENCE SNDR_TRIGGER [SNDR_LAST_LOGGED=='']/ SNDR_ND RECOVERY [SNDR_LAST_LOGGED=='INSTALLED'] SNDR_CRASH * EVENT_OK = EVENT LINK_OUTAGE WISE'06, Wuhan, China 12 T _TM means TIMEOUT

13 Outline Problem Statement and Background Interaction Contracts (IC) Framework Contract between Web Services Contract between User & Browser Implementation & Experiments: Exactly-Once Web Service (EOS) Summary WISE'06, Wuhan, China 13

14 XIC Principles XIC sender Xcom obligations None (but should resend on timeout) XIC receiver Pcom obligations Persist interaction immediately XIC sender Pcom obligations Persist state before sending message Resend message after Pcom failure XIC receiver Xcom obligations None Xcom Pcom Typical setup: Xcom = user, Pcom = browser Some failures inherently non-maskable WISE'06, Wuhan, China 14

15 IC s for Composite Web Service CUSTOMER HTML_PROMPT BUTTON_CLICKED HTML_REPLY USER1_REQ BROWSER_INPUT <XIC_I_AC BROWSER_OUTPUT CLICK_CAPTURED WEBSRVR_REP_RCVD WEBSRVR_REQ <CIC_AC WEBSRVR_REP <CIC_AC WEBSRVR_REQ_RCVD APPSRVR2_REP_RCVD APPSRVR1_REP_RCVD APPSRVR1_REQ <CIC_AC APPSRVR2_REQ <CIC_AC APPSRVR2_REQ_RCVD APPSRVR2_REP <CIC_AC APPSRVR1_REP <CIC_AC APPSRVR1_REQ_RCVD XACT_UPDATE <TIC_AC XACT_COMMITTED BROWSER_CRASH, XACT_{USER, INTERNAL}_ABORT, BROWSER_WEBSRVR_LINK_OUTAGE WEBSERVER_CRASH, APPSERVER{1;2}_CRASH, DBSRVR_CRASH, WEB_APP{1,2}_LINK_OUTAGE, APP1_DB_LINK_OUTAGE LOCAL_FAILURES GLOBAL_FAILURES WISE'06, Wuhan, China 15

16 Outline Problem Statement and Background Interaction Contracts (IC) Framework Contract between Web Services Contract between User & Browser Implementation & Experiments: Exactly-Once Web Service (EOS) Summary WISE'06, Wuhan, China 16

17 EOS: Exactly Once Web Service EOS prototype implementation: Exactly-once semantics for composite services with Transparent EOS-enabling of Web pages by piggybacked Javascript XIC s: : Transparent browser recovery via callbacks and browser-specific XML store CIC s: : App server recovery by modified session mgt. of Apache and Zend engine Fully transparent: no changes to app code neither to PHP scripts nor to browser WISE'06, Wuhan, China 17

18 Efficiency: Message Lookup Tables Built at Pcom during normal operation Rebuilt from the log during Pcom recovery Input MLT for duplicate elimination Output MLT to track CIC progress enabling timely garbage collection OMLT of eosphp3 URI MSN CIC status 3 installed 5 stable 6 unknown 7 installed 8 installed IMLT of eosphp1 client id MSN Reply LSN eosphp IMLT of eosphp2 client id MSN Reply LSN eosphp WISE'06, Wuhan, China 18

19 Experiment Setup ebay-like auction service User settings at frontend (private) Auction items at backend (shared) 5 concurrent end users, synthetic load Frontend Server P4 3Ghz, 1GB Backend Server P4 3Ghz, 1GB POST (ICIC) action=increment POST (ICIC) action=increment b2b=true Web Client <html> <p>private Count: : 3 <p>shared Count: : 1235 </html> private private private private count count count count shared count WISE'06, Wuhan, China 19

20 PHP Scripts (not changed at all!) <html> Script called Web 5 Web times Web Web 1. <html> Other server Client reports: Client Client Script Client called 1000 times 2. </html> <?php 3. session_start(); 4. $HTTP_SESSION_VARS["count"]++; 5. printf("script called %i times", 6. $HTTP_SESSION_VARS["count"]); Zend Engine 7. $ch = curl_init(" 8. $b2b_reply = Session curl_exec($ch CURL $ch); 9. printf("other server reports: %s", $b2b_reply); 10. curl_close($ch $ch); 11.?> 12.</html> Zend Engine Zend Engine Session CURL Session CURL WISE'06, Wuhan, China 20

21 Run-Time Overhead Session PHP elapsed time [sec] Web Client POST (ICIC) action=increment EOS-PHP elapsed time [sec] Overhead (elapsed time) [%] <html> <p>privatel Count: 3 <p>shared Count: 1235 </html> PHP frontend CPU time [sec] EOS-PHP frontend CPU time [sec] Frontend Server 101% private private private private count count count count step POST (ICIC) action=increment b2b=true % 5 steps steps Backend Server 93% shared count Overhead (frontend CPU) [%] 109% 122% 102% PHP backend CPU time [sec] EOS-PHP backend CPU time [sec] Overhead (backend CPU) [%] 44% 36% 33% WISE'06, Wuhan, China 21

22 Outline Problem Statement and Background Interaction Contracts (IC) Framework Contract between Web Services Contract between User & Browser Implementation & Experiments: Exactly-Once Web Service (EOS) Summary WISE'06, Wuhan, China 22

23 Summary Generic IC s s for composable services with exactly-once execution guarantee Eases programming by masking all (but inevitably non-maskable) ) failures Formal specification by state/activity- charts & model-checked CTL properties Efficient implementation: EOS prototype for Apache/PHP WISE'06, Wuhan, China 23

24 WISE'06, Wuhan, China 24